You'd be surprised how many companies are uncomfortable with full interviews with engineers. Executives are fine, but engineers can be more 'loose' with answers and say something they shouldn't. Also if an engineer does a very good interview, they might be highlighted to competitors, and the company doesn't want the engineer poached.
Just curious: what's the deal with the trio of portraits, about a page in? I found it confusing, since the top of the article has actual photos of them, and any readers of this site or your twitter feed will have seen your pic. Also, the difference in styles is a little jarring.
I usually put the photos in for people who read the article in systems where the top carousel image doesn't appear. The number of photos of Jim and Ljubisa in the wild is quite small - Tenstorrent provided the drawn portraits as even they don't have recent headshots. I had to do something with my photo so it wasn't just the regular picture, that would look more out of place imo.
I would ask the tough questions... like what are they doing to deal with the fact that as far as consumer-desktop-grade processors.. everyone and their sisters have caught up! Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one? When are they going to completely re-architecture the whole darn thing and get me some serious single-core throughput improvement? I want 10-fold speed! Sure, my new CPU looks faster with all the cool RGB.... but the only thing that's actually made it any faster has been the increased throughput in disk I/O. Heck, put an M.2 on an old R5 and it's a smoker! C'mon guys... give me some real speed and make me want a new computer again...
"Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one?"
Well that isn't true, but if you're satisfied with your current performance and don't need huge core counts, then wait until monolithic 3D CPUs appear for consumers. Putting RAM inside the CPU should yield at least that 10-fold speed up you're looking for.
> Putting RAM inside the CPU should yield at least that 10-fold speed up you're looking for.
You can find data on RAM scaling, if you're really interested, but I think you'd be disppointed. Most tasks aren't nearly so sensitive to RAM speed, due to ever wider and deeper cache hierarchies. Integrated graphics would be the main exception.
Future iterations of Apple Silicon chips are probably what you’re looking for. Intel has runs out of steam as you said. AMD are doing well with their different structural layout but they are in the same order of magnitude as Intel with respect to power, performance and price.
Apple’s A-series of chips have been impressing for a while (and made Apple a boatload of money) but their relevance to x86 performance was well-hidden. With the M-series of chips, they’re an order of magnitude better than Intel / AMD in terms of price (estimated to be IIRC $50 each) and an order of magnitude lower power, but broadly equal or better performance. And that was only the first M1 chip which still has several limitations as implemented in hardware that came on the market last year.
The next M2 will be out in a few months. No idea if it will be another 50% jump or doubling in performance but Apple’s history is excellent in this aspect - rumours also say it will improve on some of the limitations of the M1 (increased max ram, more GPU cores etc).
It all links back to what Jim was saying in the interview - full stack control, integrated hardware and software design.
> they’re an order of magnitude better than Intel / AMD in terms of price (estimated to be IIRC $50 each)
Doesn't matter if they are, when you have to pay Apple's premium prices to get them.
> an order of magnitude lower power, but broadly equal or better performance.
That's slightly misleading. You're looking only at single-threaded performance. If you look at SPEC2017 Rate-N, the 25W Mac Mini's int score comes in right between the 15 W Ryzen 4800U and the 35 W 4900 HS.
> Apple’s history is excellent in this aspect
It's not flawless. The A13's review concluded: "At the end of the day, power efficiency at the peak performance states even saw a slight regression."
Yes, even the mighty Apple isn't monotonically improving in all respects, every generation.
I would say you shouldn't expect Apple to pull off a miracle, with the M2. What they achieved with the M1 was both a new core + transitioning from a phone to a more laptop-oriented SoC. IPC improvements aren't a bottomless well, and the SoC-level changes were mostly a one-time improvement. Numbers like 50% or doubling are only going to happen with 50% or double the number of cores, and then only for multithreaded workloads.
Wake-up--it's a multi-core, multi-processing world that will never revert to single-core dependence or programming paradigms. What you seem to lack is the software that will challenge the best CPUs and GPUs today--but you obviously don't have the hardware required, so it's not surprising that you don't own the software. I would advise you not to hold your breath waiting on the tech to jump back to 20 years ago...;)
"Wake-up--it's a multi-core, multi-processing world that will never revert to single-core dependence or programming paradigms."
all the parallel-seeking software is, mostly, pie-in-the-sky. yes, multi-processing has been around since at least OS/MFT, but that's not really an application demand; each application is nearly entirely linear as it always. running more than one such on m-c, m-p cpu is an OS exercise. if, and until, someone figures out how to turn inherently linear problems into parallel ones, nothing is going to change.
a larger target, and easier to attain, is one where we return to the thrilling days of yesteryear, when cpu and memory speeds were within an order of magnitude of each other. then 3D Xpoint-like devices would make a slimmed-down cpu without all those caches and buffers, that seemingly endless 'memory hierarchy', irrelevant. all reads and writes are to durable storage. will require rethinking ACID semantics.
> each application is nearly entirely linear as it always.
Browsers have gotten a lot more parallelized, I think mostly to take advantage of multi-core mobile CPUs. Also, a lot more of the rendering has gotten pushed off to GPUs.
Games have gotten massively more threaded, in the past 10-15 years.
In fact, I think there are probably few examples of performance-intensive apps that *haven't* gone multi-threaded.
> if, and until, someone figures out how to turn inherently linear > problems into parallel ones, nothing is going to change.
The thing is, I'm not sure most people spend time running apps that are dominated by a single sequence of serial computations. If you have examples in mind, please share.
> 3D Xpoint-like devices would make a slimmed-down cpu without all those caches and buffers
No, 3D XPoint was never about getting rid of CPU caches. It's about 1/10th as fast as DRAM, at best.
At some point far in the future, we'll probably have virtual machines running on multicore processors, doing massive speculative execution and all that in software to emulate old x86. By that time, there'll probably be no software left that matters.
"Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one? "
there was a time when single-user or even networked (read: Novell) PCs ran things like good sized databases or massive spreadsheets. back then, the whole machine was taxed. with so much compute being off-loaded elsewhere, about the only use-case for more machine is uber-gamers. back then, the demand for compute was driven by The Corporations who spent like drunken sailors of PCs; and there were hordes of them. now, uber-gamers? I'm gonna guess not nearly the same TAM.
It is not a stupid claim. For many years - you buy a new PC it is definitely much faster than your old one, yet it is slower? How? Becauze Windows! No matter how faster CPU got, Microsoft managed to make them feel slower.
Software got easier to use. I remember when just saving a file took three steps and a complicated key-code. (Fire up Vim or Emacs if you want to experience that joy. (Many professionals still use them for their power, but they’re exactly accessible to the common person))
Now everyone uses cloud computing all the time. Emails docs save states etc are synced between phone and work machine and home machine. Normal non-technical people execute astonishingly complex processes on a daily basis that in the old days would have required a team of consultants digging through a stack of manuals. All that complexity is handled by layer upon layer of software now.
I think a lot of it has to do with the cumulative effects of security mitigations and multiple layers of antivirus and anti-malware. At least, that's my assessment of my work PC, but I don't know exactly what they have installed on there.
"All that complexity is handled by layer upon layer of software now"
True. In agreement with your comment. It's fantastic what software can do today. Having said that, there's also a trend towards languages that are supposedly easier, cleaner, and more secure. And what you gain on one side, you lose on the other.
2000's could mean the decade of roughly 2000-2009, which would include Nehalem.
Even Skylake is substantially faster than Nehalem. Once you factor the added cores & clockspeed of Coffee Lake-R or Comet Lake, it's game over.
And if it's IPC you want, then Sunny Cove should be Intel's reference point for the next decade, since that's when it launched in the original Ice Lake laptop chips.
I would say that CPUs got the critical mass of gains, architecture-wise, in the 90s and 2000s. Since then, it has advanced but a good deal slower than of old. That's not to say some massive increase is out of the question. It could be right round the corner. But right now it seems more like an asymptotic trailing off into infinity.
A lot's been written about the demise of Moore's Law and Dennard scaling.
In the 1990's they got free clockspeed wins from process node improvements that were actually quite large. Over the decade, clockspeeds increased by like 40x. Also, there were still a lot of instructions that were micro-coded. So, there were wins to be had by replacing it with hardwired logic, as transistor budgets increased. Then, we got on-die L1 cache, pipelining, superscalar, OoO, and vector instructions (i.e. SIMD+). The 90's ended with a ~1 Ghz P3 featuring 128-bit SSE. I/O was EIDE, PCI, AGP 2x, and USB 1.1.
The next decade was about going wider, but clockspeeds only increased by about 3.5x. Pentium 4 was designed to scale up to 10 GHz, which is why it remained fairly narrow and had so many pipeline stages. Once Intel figured out that clocks weren't going to increase like the previous decade, they course-corrected quite effectively. Core counts went up to about 4, and we got hyperthreading. Memory controllers moved on-die and L3 was added to the hierarchy. RAM went from like regular SDRAM to DDR3. Vector instructions stayed at 128-bit, but Intel fleshed out SSE and built CPUs that implemented it at full width.
If the decade we just ended had a theme, it would probably be multi-core, since that's the parameter that increased the most. Of course, CPUs continued to get wider, but clock speeds only increased by like 1.5x. Vector widths increased by 2-4x.
All of that is to say that the 1990's were about picking the last of the low-hanging fruit. Once we got into the next decade, each increment of IPC increase needed nonlinear amounts additional die area and complexity. This also increased the attack surface, leading to a bump in the road due to the need for various mitigations and more focus on designing for security.
People pretty quickly figured out that the most area-efficient way to add performance was by increasing the core count. Once software really got onboard with the effort, that is.
No matter how you slice it, the gains of the past decade were real and substantial. However, it's also true that the gains of each decade have been smaller than the previous, especially if we focus just on single-thread performance.
You summed it up excellently! I will say that going from one iteration to the next doesn't seem like much, but over time, it adds up. Even Zen to Zen 3, a mere four-year difference, is something like double the performance. Thankfully, reliability has gone up too.
I don't know much on the topic, and am very sceptical of it, but it's possible that quantum computers, once the decoherence problem is solved, could throw open the doors.
Isn't their applicability limited to a very narrow set of specialized problems? None of the quantum computers I've been reading about would ever be suitable for typical PC workloads, even leaving aside the extreme temperatures and RFI isolation they require.
Usually, encryption and searching problems. But apparently, they are Turing equivalent and can solve any problem a classical computer can---I suppose by way of iteration till they converge on the classical answer. I think current QCs are, relatively speaking, at Eniac's level. And you know, Eniac to Galaxy S10 is quite a leap, so QCs might have some way to go still. I'm a bit sceptical of the field though and wonder if there truly is any difference between a QC and a probabilistic classical computer.
Turing-equivalent just means that you can map the set of classical computation problems onto them, without regard for speed or space. It doesn't make them *good* at implementing classical computation, in a similar way that a lot of classical algorithms run poorly on something like a GPU.
I think we'll probably still rely on conventional CPUs to implement conventional algorithms and just use quantum computers for large-scale, multi-parameter optimization problems. It doesn't mean they won't be revolutionary, just that they're not going to improve all of the things that faster conventional CPUs do.
Perhaps in the far future, once they've been shrunk enough and actually work, one could have a CPU that carries a little quantum unit. It could help with the tasks QCs excel at.
It's an idea, but entanglement seems far too fragile to do without enclosures on the caliber used for experiments on particle physics. Maybe I'll be proven wrong, but I expect the only way that 99.999% of the population ever encounters quantum computers is via the cloud.
Anyway, we still need to solve the scaling problems, so they actually provide real benefits, before it's even worth focusing on ways to make them smaller, cheaper and remotely practical for an average person to have in their home.
You're right. Not worth it as it stands. I think a probabilistic classical computer could be a more feasible approach. How close it would approach a QC, I don't know, but something like that could be up and running in no time. (If you can't get real-time lighting, go for lightmaps.) As for QCs, it's almost as if Nature is reluctant to co-operate. Some hint?
I think a deeper insight, coming from theory, is also lacking and might lead to a better understanding of what exactly is going on at the lowest level. Clearly, superposition, entanglement, etc. are springing from some truth that's been caught in paradoxical classical terms. The secret lies with entanglement, I'd say; and concerning the other quantum mysteries, how much of them are spectres of the mathematics or physical reality.
Anyhow, there are many itching to get their hands on quantum computers because of the implications for security, and they'll chase it to the ends of the earth. Emperor's New Clothes? Only time will tell.
> there are many itching to get their hands on quantum computers because of the implications for security
Not only that, but there are lots of high-value optimization problems. I see references to things like material science and drug discovery. D-Wave is also talking about deep learning, although that would have to be a very small model to train on their current-gen systems, I think.
Lots of biological systems, even plants, make use of quantum features, so these sort of computers could be a better fit for simulating them. I expect it would aid physics too, especially with simulating the quantum side of black holes.
Except that's not a tough question. For these two guys it's about the most irrelevant question ever. They are not in the consumer-desktop processor business so how in the wide world of sports does that have anything to do with them?
Yeah, I thought about making that point. I guess it's tangentially relevant for Jim, since he just left Intel about 1 year ago, presumably working on consumer-desktop processors (among others). But, it really had nothing to do with the focus of the article.
Anyway, the thread had already taken off, by then. I figured I'd play along.
Great article, cool to hear about a company starting in the a literal basement of Toronto (shows you how crazy their real estate prices are he still can't get out of his mom's basement). Kinda sad to hear about the internal management of AMD and Intel being like this (it sounds like they're hiring suits rather than real engineers for upper management which is a big red flag):
AMD: I wasn’t a Senior Fellow, I was actually a director - everybody at the time kept saying that nobody understands why I’m only a director and why am I not a fellow, or a senior fellow. That was a common theme there, but I guess I fit in more with these technical folks and they said there are a lot of organizational challenges to getting anything serious done. I thought that it was better [for me to be] positioned somewhere you have a bit of reach into both.
Intel: Intel, people were surprised a Senior VP was grilling people on how they manage their batch queue, or what the PDK is, or what the qualification steps were, or how long the CAD flows took to run, or what the dense utilization was in the CPU, or how the register file was built. I know details about everything, and you know I care about them, I actually really care. I want them to look good. A friend of mine said, if it doesn't look good, it isn’t good. So when you look at the layout of a processor, or how something is built, it's got to be great. You can't make it great if you don't know what you're doing. It’s bad if somebody gives you like a five-level abstraction PowerPoint slide that said, things are going good, and we're improving things by 3% per year - I've been in meetings where somebody said 5% better, and I said ‘better than what? Better than last year, or better than the best possible, or better than your first-grade project?’. They literally didn't know what - they'd been putting 5% better on the slide for so long, they forgot what it was.
While obviously Bulldozer is the highlight "problem," worth remembering that AMD spent 3 generations on 32nm and a 4th on 28nm bulk. Bulldozer wasn't just one bad design cycle.
Guy, The President of AMD is an MIT engineer...;) AMD is filthy with engineers. It wasn't suits who put the company where it is today and it won't be suits that will keep them there.
> it sounds like they're hiring suits rather than real engineers for upper management which is a big red flag
So, um, maybe you didn't hear about this, but both Intel and AMD are now run by engineers.
As for your quote of Ljubisa's, that seemed pre-Zen. Obviously, AMD had issues at that time.
Jim didn't say the second part of your quote was specifically at Intel. If anything it sounded more like the beginning of his second stint and AMD, but he could've just been taking those examples from other presentations he sat through, in his long career.
Fantastic interview! Only halfway through but was grateful for the remarks on Zen's development; and I see there's a whole load of nutritional goodness up ahead. RISC-V, the works. Can't wait to read the rest (but eyes are sadly closing this evening).
The light it shed on the development of Zen and basically AMD's whole transformation was nothing short of amazing! This interview will quite likely be cited by tech historians chronicling that era.
I think it's a fine line to walk, when trying to push people to step up and do better work. You mostly have to challenge, while being supportive. And you don't want to ruffle too many feathers and make enemies, or compromise your vision for the sake of consensus. But, you also don't want to seem impossibly demanding, so you need to bring people along towards your vision and not just cram it down their throats.
I think that's why having a team of experts who can & have all walked the walk is so powerful. Because, you can really challenge anyone who tries to throw up roadblocks, make excuses, or dig in their heels. And in tearing down opposition, you can also impress on everyone else what can be accomplished by a relentless and creative pursuit of solutions (instead of excuses).
*sigh*. It makes me wish I could work on a team like that. Sadly, the only inspiring leaders I've worked under were borderline jerks. And their plans tended to be half-baked, at best.
Yes, that was the golden part for me. The drawing back of the curtains, if only for a moment. From those few paragraphs, I can imagine the spirit of despair that had taken hold of AMD, along with their lack of belief. I'm glad Jim came in and shook them all up. And the right sort of man to put those geniuses in their places who were trying to block Zen. I think it's fair to say they'd have been content carrying on with Bulldozer.
"Sadly, the only inspiring leaders I've worked under were borderline jerks."
Their personalities tend to be touched with an autocratic, self-centered ruthlessness, which, unfortunately, can lead to success but hurts many along the way. (I've read Steve Jobs was like that.) Funny thing is, folk fall at such characters' feet. Political/military leaders like this are famous.
In my case, they had a vision and a passion that would inspire others, but tended to be dismissive of people who didn't agree and could be rather petty an undiplomatic. I think that's kind of the norm, for the tech industry.
It takes some patience and energy to bring people along with you, but sometimes they ask good questions or have good ideas, in the process. Plus, you can get more people on board and pulling in the same direction than with a sharp-elbowed approach. So, I think the key is to see that someone is digging in their heels and then to find a good way around them without too much fuss.
People seem to think a good leader has to be like Steve Jobs and break a few eggs to make a good omelet. I think a more avuncular approach is probably better, though it shouldn't be confused with being obsequious. You want to challenge people, in a supportive manner, and give them a real chance to come along -- yet, without compromising in your goals.
Wholly agree. For my part, when a leader/manager treats you as a human, as an individual, tries to listen to what you've got to say, and considers it, whether he or she agrees or not---that's a person one desires to serve, follow, and please. I suppose it all comes down to their manner, even the tone of voice, and their being humble and down-to-earth. It comes naturally to some. Others haven't got it. A true leader does take charge and lead (without compromise on certain points), but doesn't put down others. Treats them with dignity, rather. Such a leader, one loves.
This was much better than the interview with Keller when Keller was at Intel. My guess is that for the Intel interview Keller had instructions from Intel's PR department on what to say and what not to say. So he was quite limited in what he could say about what he was doing at Intel, and he was no doubt prohibited from saying anything whatsoever about his work at AMD. It's also possible that he didn't like working at Intel very much.
In this interview, he's talking pretty freely about stuff he's passionate about.
From the present interview, I get the impression that, at Intel, there's a lack of proper communication between the different levels (along with too big teams). Perhaps that's partly why they're in the state they are today. Hopefully, Mr. Gelsinger can stir things up there.
All the rumors i've heard was that he essentially left Intel because they wouldn't do the radical changes he wanted, so my guess is that anything Intel releases that was designed by him would be underwhelming.
In the comments thread soliciting questions for the next interview segment, Ian confirmed that Jim left for family reasons (and yes, *real* family reasons, even though that's the usual line people give). Unfortunately, it got deleted after the interview happened.
However, if you read the article, he had put his own money into Tenstorrent, before he even joined Intel. And after hearing about how many doors were opened to them after he started working at Tenstorrent, I think he really didn't need any other reason to go there than to help them succeed.
Plus, he's notorious for leaving places after only a few years. His stint at Intel was short, but not totally out of character for him.
Excavator wasn't awful either lol, it had way better power characteristics than any prior Bulldozer derivative and did quite well when it first came out, but its biggest benefits were platform changes for AM4 that it didn't utilize all that well, but its successors did.
This is also a good description of Nehalem vs Sandy Bridge. Nehalem's biggest changes were platform-level in migrating toward an SoC-like design. Its performance benefits were nice over the previous generation but the big overhaul of Sandy Bridge is what people remember for a reason -- they were more important.
The relative performance positions of AMD vs Intel in those comparisons is the primary reason you're objecting imo. Doesn't change the internal jump, which is my point. The wild ride of micro-op caches cannot be understated. Neither can the jump from the weird "permanent SMT" Bulldozer module to Zen's CCX.
Thanks for sneaking that silicon photonics question in. I suppose he's right, tech "enthusiasts" are always talking about the next big thing that will revolutionize everything while the foundries are busy actually improving what already works, and people like Jim are re-thinking how to use the different tools to actually do what they need being done.
Still, while I find his enthusiasm for technology great, I still can't but be skeptical of AI. At some point they'll be able to tell which kind of books you might be reading just by the words you use or the way you structure your sentences. Imagine a totalitarian dictatorship having that kind of power. Perhaps is the kind of power that humans don't deserve to wield.
Hopefully, the second part of the interview will go into the ethics and other concerns around AI. As we can appreciate that Jim is a thoughtful guy, I'd expect he's contemplated it in some depth.
> At some point they'll be able to tell which kind of books you might be reading just by the words you use or the way you structure your sentences.
They already can. Check out OpenAI’s GPT-2. You can feed it a text snippet and it will write an article using the same phrasing and topic style. I believe a similar or later version had to be withdrawn from general release because it was too realistic and made it too easy to mass-produce political propaganda.
Thank you for the interview. I’ve really come to appreciate what makes Jim Keller so great after reading/watching some of his interviews. Working in the tech industry for awhile now and working for what has become a mega-corporation, I can relate to the things he says and some of his comments about failing cultures. It really does go beyond the technology. It’s kind of scary listening to the AMD story and knowing that most corporations will ride failing leadership/culture to the bottom until there is enough desperation to change things. And most times you don’t have a JK to save you. I wish I had the chops to be part of his circles.
It's really nice for tenstorrent to make their accelerator cards available to the public. Availability of hardware is crucial to reproducibility of AI research. Looking forward to getting one!
If he said exactly that, I must've missed it. However, their website says they are already bringing a devcloud online, and it currently supports inferencing workloads. They also have a link to apply for early access to Greyskull.
That's actually quite cool! I mean, according to their investment rounds, they are burning through the millions. And still [coming soon] you can simply buy a card for 1000$, download the software for free and have a go with it! That's só cool!
I don't really get the flight booking situation that LB mentioned. Did he accidentally book a flight one week earlier than on a date he was supposed to travel?
Yeah, I wish that had been a little clearer, on a couple points. I'm inferring that Jim's advice to him was to buy a ticket on his own dime, because missing those meetings would do too much damage to his standing in the organization. But, it'd be nice if it clarified exactly what Jim was referring to and what Ljubisa ultimately did about it.
That said, it was a tangent, and its purpose was really just to show how Jim would give him career advice. So, we needn't worry too much about the particulars.
Fascinating. Keller comes across a a breath of fresh air, like someone who wants to change the (silicon) world. Seems like he has a ton of good energy. From how Keller describes intel and the bureaucracy, it sounds like it was killing him slowly inside. Hopefully he left a mark and we see good things from the fruit of his labor, similar to what we saw at AMD.
I think Keller's fine--like the guy just fine. But at some point people need to get past the "hero" syndrome, imo...;) Playing musical chairs with your employment is fine when you have the necessary funds under your belt to ensure that you don't starve no matter what happens. Most company employees don't have that privilege. But getting to where AMD is today is 100% a collaborative effort undertaken by thousands of employees and engineers all contributing something of importance. Keller's now with another company so naturally he will promote himself and his new company while he criticizes the other guys...it's just business 101. What I really, really like about AMD is that the company isn't held together by "rockstar" engineers and programmers--indeed, teamwork happens to be at a premium at AMD, imo. Intel, of course, is pretty much a lost cause--the company is so waterlogged and rotten with turf wars and monopoly thinking it's a miracle they can produce much of anything competitive. The company had to bring an old man out of retirement because nobody there had an idea of what to do next regardless of how much money they were drawing down. Really pitiful, imo.
I don’t think JK would say AMD’s success was due to one person. He’s always come off as an anti-prima donna/blue collar kind of guy to me. But the reality is that for every really good engineer with great ideas, there are 2 or 3 really bad engineers. And the bad ones would rather fall on their sword than ever admit they are wrong. If you don’t have guys like Keller to 1) correctly sort good and bad ideas 2) find and enable good engineers 3) check/block/remove bad engineers by any means necessary 4) set the direction and expectations for execution, your company will under perform. Sounds easy but it’s amazing how hard it is to find those X-factor guys. So why should the average employee care? Because in the commercial world it makes the difference between growth/bonuses and you losing your job. If you’re working in academia or defense, maybe you don’t care. But a lot of people appreciate a track record of delivering good products.
> Our initial target isn't to get some huge contract, it is to get 100 programmers > using our hardware, programming it, and living with it every day.
Sadly, AMD still doesn't get this. Nvidia got it, and it helped give them their lead in machine learning. But AMD's Linux driver stack for GPU-compute has been absent or broken for RDNA cards, since day one.
All AMD's GPU-Compute division seems to care about is HPC and hyperscalers. They don't seem to understand that it's kids in university who are pioneering the new frameworks and algorithms. And they're using gaming cards, not $10k compute accelerators!
AMD's first priority should've been top-to-bottom hardware support, for their entire software stack! Even if it meant missing out on a couple supercomputer wins, the long-term payoff would outweigh that by orders of magnitude.
Not too long ago, a basic thing like hardware encoding was broken on AMD GPUs in Handbrake. I remember the HB team contacted AMD, and it was fixed after that, requiring updates to both the driver and HB.
I got from this that they don’t yet have more than 15 people in the world able to program their hardware. Still very early days, when counting everyone who knows how code for your new stack just means taking off your socks.
> they don’t yet have more than 15 people in the world able to program their hardware.
I think his point was that he was drawing a contrast with other AI startups, who just want to deliver their solution as a black box that customers only use via supported deep learning frameworks.
Tenstorrent wants people to open the box and tinker inside. Based on their website, they seem interested for people to use their chips even for non-deep learning applications.
> LB: You get the flexibility, definitely. Adding programmability isn’t free as usual, > so there’s a bit of sacrifice there.
This is why I've long thought GPUs had an advantage over a lot of the AI accelerators. GPUs give up a bit of performance, but they offer a lot in programmability. And they're very mature, in that respect. There's obviously code that runs better and code that runs worse on them, but they're still quite versatile, compared with a lot of what's out there, and you can usually find a way to still get decent performance.
the move to seperate compute and gaming architectures means they have very little reason on porting their new software stack to old gaming cards, I'm afraid.
You're replying to my above point about RDNA? If AMD thinks they don't care about RDNA as a viable platform for compute, then they're basically saying they want AMD GPUs to be about as relevant to the future of computing as the other HPC-exclusive players you've probably never heard of.
Nvidia gives away full CUDA support, on their *entire* hardware line, from $100 Jetson boards on up. Sure, some of their GPUs have stuff a bit nerfed, like reduced NVENC streams and fp16 tensors running at half speed, but it's still usually fast enough to be worthwhile, and certainly you can use it as a software development platform.
JK> We are trying to meet the software guys where they are at, because they JK> just want to write code. They also want to understand the hardware JK> so they can drive it, but they don't want to be tortured by it.
So true. You want hardware that's interesting enough to be exotic and to let you push the envelope, but not so byzantine that you spend 98% of your time and energy catering to it, rather than focusing on your software and algorithms.
JK> I know details about everything, and you know I care about them, I actually really care.
I think that's the secret. That's how you earn the respect of the rank-and-file. You show that you understand and care about their work, while also demonstrating competence they can appreciate. It also helps keep them on their toes, letting them know that they're not going to get by with second-rate work.
JK> A friend of mine said, if it doesn't look good, it isn’t good. JK> You can't make it great if you don't know what you're doing.
Exactly. There's no room for BS, when you're reaching for excellence. You not only need to understand how something works, but how and why it's better than all the alternatives and why it can't be even better.
JK> I've been told at various points in my career to focus on that high-level picture of JK> managing, but I always like to get into the details
I've worked for managers that spent all their time in the details and didn't do enough managing. That was not good. At some point, the opportunity cost of managers not managing becomes greater than the value of the technical contributions they can make. In Jim's case, it sounds more like he was acting as a consultant, rather than trying to do the actual work of an individual contributor.
But, I've also seen middle-managers try to lead an engineering organization, without the core technical chops. They didn't know when they were being fed a load of BS, didn't hire the right people under them, and tended to make decisions more for political than technical or sound business reasons. And that's the kind of stuff that can rot a department from the inside out.
JK> abstraction stack - there's the high level, there's the medium, and there's a low level. JK> Yes, you do need to know a lot about all of them,
I'd say: pick your level and own it. Master it as well as anyone. But, you've got to know enough about the other levels that you at least know when to go there, and can pick up what else you need, as necessary.
I've known a lot of generalists, and I've got to say that they're rarely as interesting or impressive as the specialists. Just don't be afraid to go outside your domain.
They are 2 different kinds of workers. As a generalist myself, I try to stay up-to-date with all the fancy stuff the specialists are working on. I use this knowledge to integrate it with the work of other specialists. While the specialists use it to advance in their field. Of which they then again keep me up to date, so I can integrate it with the work of others. And so on.
ps. For a specialist, there are só many opportunities to advance, só many options. But there's only a limited amount of time, so they have to choose their next steps carefully. This is where people like me try to be of value; by having an overview, I can make them realise what next step could be beneficial to the company as a whole or to the platform they are working in, and which would be wáy out of line and/or very hard to integrate.
Thanks for sharing your experience. I'm aware that some places with a lot of specialists tend to pair them with generalists, who can help integrate their research into real products.
It's rare, but you sometimes see a specialist with all the versatility of a generalist.
JK> I heard from many people that Jack could get to the bottom of anything.
A lot of it just comes down to fearlessness and not being afraid to ask a dumb question or seem ignorant. And if you ask enough questions, you start to learn which are the key ones. That and persistence will get you most of the way there.
One question I have is why keep the ARC cores, once they start integrating RISC-V. I guess that's probably not yet set in stone (or lithography masks, as it were), but I'd expect them to just go RISC-V everywhere they don't have a good reason not to.
It's interesting to note their lineage traces back to the FX processors in a couple Super NES game cartridges. I'm guessing they ended up in Grayskull via its FPGA-based predecessor, as they now seem to be part of Synopsys' licensable IP library.
And, from a user's perspective, I'm wondering a lot about how DRAM fits into the programming model. Presumably, it can be accessed from any node in the graph, even if doing so isn't exactly fast.
I think DRAM is, among other things, key to their scaling story. It's a fundamental part of how they could take a processing graph and scale it down to a single chip, even if the entire set of weights or some intermediate state is too big to fit in the chip's SRAM. In that sense, how it's accessed could be an easily-overlooked means of differentiating themselves vs. others taking a similar sort of dataflow-oriented approach.
And speaking of their competition, I'm struck by some apparent similarities between their approach and that of Cerebras and perhaps even Habana and Graphcore. Basically, the idea of building a dataflow processor and compiling a processing graph to fit either one chip or span multiples, via high-bandwidth networking and a roughly mesh-like topology.
To the extent that description is accurate, it could be something akin to the formula the industry has settled on for building fast GPUs (i.e. wide SIMD + deep SMT), or even fast CPU cores.
I put off reading it for a week as it was so long but I’m glad it read it in the end. Not a word was wasted. Some other interviews on AnandTech have been full of corporate speak and waffle so I was reluctant to read this one. But this one really delivers. Excellent! I suspect you’ll get a lot of hits coming back to this article over the next few years.
Looking forward to the next one with Jim!
PS I lol’d at the question: ‘what’s the limitations of your new chip?’ Just like being asked at a job interview ‘What’s your greatest weakness?’ I’ve sometimes asked that at interviews for kicks & to see how they deal with difficult questions - and how well prepared they are. The answer itself doesn’t really matter as long as it’s delivered well. Usually get something like ‘I work too hard’ or ‘I care too much about the job’. I don’t ask it any more as we sometimes get (excellent) disabled candidates and it would be wrong (and potentially discriminatory) to ask them ‘what’s your greatest weakness?’. There are other ways of asking difficult questions in interviews.
> I don’t ask it any more as we sometimes get (excellent) disabled candidates
You could use a simple variation: "What's one big decision you've made in a project that you'd now make differently?"
In some ways, it's better than your original question, since it reveals how thoughtful and invested in their projects they tend to be, as well as their sense of agency.
Jim's super cool and also very respected in the industry. He was a big inspiration for me early on in my career. I was a fairly junior CPU performance engineer at AMD when Jim joined AMD in 2012 (he'd worked at AMD earlier in his career too). I worked on the Zen program, which was being led by Jim. I saw him at my gym one morning, deadlifting 275 lbs (or maybe more) and I went up to him and introduced myself. He was super friendly and continued to be so whenever I'd run into him. We'd exchange our personal bests in lifting. IIRC, he was in his mid 50's then. Given the similarity in our backgrounds (CPU design) and his professional achievements and his amazing discipline, it was a no-brainer for me to aspire to be like him. www.ilmibook.com says This is a man at the cutting edge of his field.
Thank you for informative article .are you searching for help to get ideas about the Digital marketing service? If yes, then you should try the..<a href="https://lsgodigital.co.uk/5-ways-to-digitally-prom... Ways To Digitally Promote Your Business </a>
Thank you for informative interview .are you searching for help to get ideas about the Digital marketing service? If yes, then you should try the.<a href="https://lsgodigital.co.uk/5-ways-to-digitally-prom... Ways To Digitally Promote Your Business </a>
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
130 Comments
Back to Article
ralritt - Thursday, May 27, 2021 - link
Great Kim Jeller interview!mode_13h - Friday, May 28, 2021 - link
Absolutely! This is like a treasure trove of all kinds of details we rarely ever hear!I really have a lot more respect for Jim, after reading this. I wish nothing but the best for them.
GeoffreyA - Saturday, May 29, 2021 - link
It was great to get a glimpse of Jim's personality, and one sees he's such a likable, down-to-earth man. A lot of humility too.As for the "treasure trove," goodness gracious me, it's overflowing!
outsideloop - Thursday, May 27, 2021 - link
I would love to see an interview with Mike Clark in the future, Ian.Ian Cutress - Thursday, May 27, 2021 - link
You'd be surprised how many companies are uncomfortable with full interviews with engineers. Executives are fine, but engineers can be more 'loose' with answers and say something they shouldn't. Also if an engineer does a very good interview, they might be highlighted to competitors, and the company doesn't want the engineer poached.mode_13h - Thursday, May 27, 2021 - link
Just curious: what's the deal with the trio of portraits, about a page in? I found it confusing, since the top of the article has actual photos of them, and any readers of this site or your twitter feed will have seen your pic. Also, the difference in styles is a little jarring.No big deal, but it had me wondering.
Ian Cutress - Thursday, May 27, 2021 - link
I usually put the photos in for people who read the article in systems where the top carousel image doesn't appear. The number of photos of Jim and Ljubisa in the wild is quite small - Tenstorrent provided the drawn portraits as even they don't have recent headshots. I had to do something with my photo so it wasn't just the regular picture, that would look more out of place imo.Fataliity - Thursday, May 27, 2021 - link
What is your timeline for releasing the second half of this article? Soon? Or like a few months?Spunjji - Friday, May 28, 2021 - link
I enjoyed seeing you rendered as what appears to be an incorporeal being constructed from discrete data vectors 😅GeoffreyA - Friday, May 28, 2021 - link
Best picture. Sort of reminded me of the Dahaka from Warrior Within.Tomatotech - Saturday, June 5, 2021 - link
I appreciated the second set of labelled headshots as it let me know who was who.mode_13h - Sunday, June 6, 2021 - link
Ah, I hadn't even noticed the main image wasn't captioned, as I already knew what Jim looks like. That would be a better solution.TheyCallMeDanger - Thursday, May 27, 2021 - link
I would ask the tough questions... like what are they doing to deal with the fact that as far as consumer-desktop-grade processors.. everyone and their sisters have caught up! Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one? When are they going to completely re-architecture the whole darn thing and get me some serious single-core throughput improvement? I want 10-fold speed! Sure, my new CPU looks faster with all the cool RGB.... but the only thing that's actually made it any faster has been the increased throughput in disk I/O. Heck, put an M.2 on an old R5 and it's a smoker! C'mon guys... give me some real speed and make me want a new computer again...nandnandnand - Thursday, May 27, 2021 - link
"Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one?"Well that isn't true, but if you're satisfied with your current performance and don't need huge core counts, then wait until monolithic 3D CPUs appear for consumers. Putting RAM inside the CPU should yield at least that 10-fold speed up you're looking for.
mode_13h - Friday, May 28, 2021 - link
> Putting RAM inside the CPU should yield at least that 10-fold speed up you're looking for.You can find data on RAM scaling, if you're really interested, but I think you'd be disppointed. Most tasks aren't nearly so sensitive to RAM speed, due to ever wider and deeper cache hierarchies. Integrated graphics would be the main exception.
Tomatotech - Saturday, June 5, 2021 - link
Future iterations of Apple Silicon chips are probably what you’re looking for. Intel has runs out of steam as you said. AMD are doing well with their different structural layout but they are in the same order of magnitude as Intel with respect to power, performance and price.Apple’s A-series of chips have been impressing for a while (and made Apple a boatload of money) but their relevance to x86 performance was well-hidden. With the M-series of chips, they’re an order of magnitude better than Intel / AMD in terms of price (estimated to be IIRC $50 each) and an order of magnitude lower power, but broadly equal or better performance. And that was only the first M1 chip which still has several limitations as implemented in hardware that came on the market last year.
The next M2 will be out in a few months. No idea if it will be another 50% jump or doubling in performance but Apple’s history is excellent in this aspect - rumours also say it will improve on some of the limitations of the M1 (increased max ram, more GPU cores etc).
It all links back to what Jim was saying in the interview - full stack control, integrated hardware and software design.
Tomatotech - Saturday, June 5, 2021 - link
I intended to reply to TheyCallMeDanger oops.mode_13h - Sunday, June 6, 2021 - link
> they’re an order of magnitude better than Intel / AMD in terms of price (estimated to be IIRC $50 each)Doesn't matter if they are, when you have to pay Apple's premium prices to get them.
> an order of magnitude lower power, but broadly equal or better performance.
That's slightly misleading. You're looking only at single-threaded performance. If you look at SPEC2017 Rate-N, the 25W Mac Mini's int score comes in right between the 15 W Ryzen 4800U and the 35 W 4900 HS.
> Apple’s history is excellent in this aspect
It's not flawless. The A13's review concluded: "At the end of the day, power efficiency at the peak performance states even saw a slight regression."
Yes, even the mighty Apple isn't monotonically improving in all respects, every generation.
I would say you shouldn't expect Apple to pull off a miracle, with the M2. What they achieved with the M1 was both a new core + transitioning from a phone to a more laptop-oriented SoC. IPC improvements aren't a bottomless well, and the SoC-level changes were mostly a one-time improvement. Numbers like 50% or doubling are only going to happen with 50% or double the number of cores, and then only for multithreaded workloads.
WaltC - Friday, May 28, 2021 - link
Wake-up--it's a multi-core, multi-processing world that will never revert to single-core dependence or programming paradigms. What you seem to lack is the software that will challenge the best CPUs and GPUs today--but you obviously don't have the hardware required, so it's not surprising that you don't own the software. I would advise you not to hold your breath waiting on the tech to jump back to 20 years ago...;)FunBunny2 - Friday, May 28, 2021 - link
"Wake-up--it's a multi-core, multi-processing world that will never revert to single-core dependence or programming paradigms."all the parallel-seeking software is, mostly, pie-in-the-sky. yes, multi-processing has been around since at least OS/MFT, but that's not really an application demand; each application is nearly entirely linear as it always. running more than one such on m-c, m-p cpu is an OS exercise. if, and until, someone figures out how to turn inherently linear problems into parallel ones, nothing is going to change.
a larger target, and easier to attain, is one where we return to the thrilling days of yesteryear, when cpu and memory speeds were within an order of magnitude of each other. then 3D Xpoint-like devices would make a slimmed-down cpu without all those caches and buffers, that seemingly endless 'memory hierarchy', irrelevant. all reads and writes are to durable storage. will require rethinking ACID semantics.
FunBunny2 - Friday, May 28, 2021 - link
"that seemingly endless 'memory hierarchy', irrelevant"that seemingly endless 'memory hierarchy', quite relevant
we need and edit.
mode_13h - Friday, May 28, 2021 - link
> each application is nearly entirely linear as it always.Browsers have gotten a lot more parallelized, I think mostly to take advantage of multi-core mobile CPUs. Also, a lot more of the rendering has gotten pushed off to GPUs.
Games have gotten massively more threaded, in the past 10-15 years.
In fact, I think there are probably few examples of performance-intensive apps that *haven't* gone multi-threaded.
> if, and until, someone figures out how to turn inherently linear
> problems into parallel ones, nothing is going to change.
The thing is, I'm not sure most people spend time running apps that are dominated by a single sequence of serial computations. If you have examples in mind, please share.
> 3D Xpoint-like devices would make a slimmed-down cpu without all those caches and buffers
No, 3D XPoint was never about getting rid of CPU caches. It's about 1/10th as fast as DRAM, at best.
shadowjk - Tuesday, June 1, 2021 - link
At some point far in the future, we'll probably have virtual machines running on multicore processors, doing massive speculative execution and all that in software to emulate old x86. By that time, there'll probably be no software left that matters.FunBunny2 - Friday, May 28, 2021 - link
"Why is my old moldy vintage 2000's CPU just as fast as my brand spanking new one? "there was a time when single-user or even networked (read: Novell) PCs ran things like good sized databases or massive spreadsheets. back then, the whole machine was taxed. with so much compute being off-loaded elsewhere, about the only use-case for more machine is uber-gamers. back then, the demand for compute was driven by The Corporations who spent like drunken sailors of PCs; and there were hordes of them. now, uber-gamers? I'm gonna guess not nearly the same TAM.
Kvaern1 - Saturday, May 29, 2021 - link
Your old 2000's CPU would be a 1 GHZ CPU with about half the IPC of current CPU'sDon't make stupid claims.
Zingam - Sunday, May 30, 2021 - link
It is not a stupid claim. For many years - you buy a new PC it is definitely much faster than your old one, yet it is slower? How? Becauze Windows! No matter how faster CPU got, Microsoft managed to make them feel slower.GeoffreyA - Sunday, May 30, 2021 - link
Got to agree. CPUs get faster, and software covers the ground in no time.Tomatotech - Saturday, June 5, 2021 - link
Software got easier to use. I remember when just saving a file took three steps and a complicated key-code. (Fire up Vim or Emacs if you want to experience that joy. (Many professionals still use them for their power, but they’re exactly accessible to the common person))Now everyone uses cloud computing all the time. Emails docs save states etc are synced between phone and work machine and home machine. Normal non-technical people execute astonishingly complex processes on a daily basis that in the old days would have required a team of consultants digging through a stack of manuals. All that complexity is handled by layer upon layer of software now.
Tomatotech - Saturday, June 5, 2021 - link
* not exactly accessible to the common personmode_13h - Sunday, June 6, 2021 - link
I think a lot of it has to do with the cumulative effects of security mitigations and multiple layers of antivirus and anti-malware. At least, that's my assessment of my work PC, but I don't know exactly what they have installed on there.GeoffreyA - Monday, June 7, 2021 - link
"multiple layers of antivirus and anti-malware"Disabling my AVG antivirus does cause the computer to go slightly faster. And there's an AVG disk driver attached to the drives. So, more overhead.
GeoffreyA - Monday, June 7, 2021 - link
"All that complexity is handled by layer upon layer of software now"True. In agreement with your comment. It's fantastic what software can do today. Having said that, there's also a trend towards languages that are supposedly easier, cleaner, and more secure. And what you gain on one side, you lose on the other.
mode_13h - Sunday, May 30, 2021 - link
2000's could mean the decade of roughly 2000-2009, which would include Nehalem.Even Skylake is substantially faster than Nehalem. Once you factor the added cores & clockspeed of Coffee Lake-R or Comet Lake, it's game over.
And if it's IPC you want, then Sunny Cove should be Intel's reference point for the next decade, since that's when it launched in the original Ice Lake laptop chips.
mode_13h - Sunday, May 30, 2021 - link
And I forgot AVX. That didn't launch until like 2011, in Sandybridge.PCs of the 2000's were also still on SATA2 and PCIe 2.0.
GeoffreyA - Sunday, May 30, 2021 - link
I would say that CPUs got the critical mass of gains, architecture-wise, in the 90s and 2000s. Since then, it has advanced but a good deal slower than of old. That's not to say some massive increase is out of the question. It could be right round the corner. But right now it seems more like an asymptotic trailing off into infinity.mode_13h - Sunday, May 30, 2021 - link
A lot's been written about the demise of Moore's Law and Dennard scaling.In the 1990's they got free clockspeed wins from process node improvements that were actually quite large. Over the decade, clockspeeds increased by like 40x. Also, there were still a lot of instructions that were micro-coded. So, there were wins to be had by replacing it with hardwired logic, as transistor budgets increased. Then, we got on-die L1 cache, pipelining, superscalar, OoO, and vector instructions (i.e. SIMD+). The 90's ended with a ~1 Ghz P3 featuring 128-bit SSE. I/O was EIDE, PCI, AGP 2x, and USB 1.1.
The next decade was about going wider, but clockspeeds only increased by about 3.5x. Pentium 4 was designed to scale up to 10 GHz, which is why it remained fairly narrow and had so many pipeline stages. Once Intel figured out that clocks weren't going to increase like the previous decade, they course-corrected quite effectively. Core counts went up to about 4, and we got hyperthreading. Memory controllers moved on-die and L3 was added to the hierarchy. RAM went from like regular SDRAM to DDR3. Vector instructions stayed at 128-bit, but Intel fleshed out SSE and built CPUs that implemented it at full width.
If the decade we just ended had a theme, it would probably be multi-core, since that's the parameter that increased the most. Of course, CPUs continued to get wider, but clock speeds only increased by like 1.5x. Vector widths increased by 2-4x.
mode_13h - Sunday, May 30, 2021 - link
All of that is to say that the 1990's were about picking the last of the low-hanging fruit. Once we got into the next decade, each increment of IPC increase needed nonlinear amounts additional die area and complexity. This also increased the attack surface, leading to a bump in the road due to the need for various mitigations and more focus on designing for security.People pretty quickly figured out that the most area-efficient way to add performance was by increasing the core count. Once software really got onboard with the effort, that is.
No matter how you slice it, the gains of the past decade were real and substantial. However, it's also true that the gains of each decade have been smaller than the previous, especially if we focus just on single-thread performance.
GeoffreyA - Sunday, May 30, 2021 - link
You summed it up excellently! I will say that going from one iteration to the next doesn't seem like much, but over time, it adds up. Even Zen to Zen 3, a mere four-year difference, is something like double the performance. Thankfully, reliability has gone up too.GeoffreyA - Sunday, May 30, 2021 - link
* Reliability, compared to the early 2000s.mode_13h - Sunday, May 30, 2021 - link
Speaking of Moore's Law, here's a recent talk Jim Keller gave on the subject (haven't watched, but I plan to):https://www.youtube.com/watch?v=oIG9ztQw2Gc
"Jim Keller: Moore’s Law is Not Dead"
UC Berkeley
EECS Colloquium
Wednesday, September 18, 2019
mode_13h - Sunday, May 30, 2021 - link
And then there's this, by Ljubisa Bajic and Jasmina Vasiljevic at Hot Chips Symposium 2020.https://www.youtube.com/watch?v=HLjumOyWj0g
"Machine Learning vs. Moores Law"
GeoffreyA - Sunday, May 30, 2021 - link
I don't know much on the topic, and am very sceptical of it, but it's possible that quantum computers, once the decoherence problem is solved, could throw open the doors.mode_13h - Sunday, May 30, 2021 - link
Isn't their applicability limited to a very narrow set of specialized problems? None of the quantum computers I've been reading about would ever be suitable for typical PC workloads, even leaving aside the extreme temperatures and RFI isolation they require.GeoffreyA - Sunday, May 30, 2021 - link
Usually, encryption and searching problems. But apparently, they are Turing equivalent and can solve any problem a classical computer can---I suppose by way of iteration till they converge on the classical answer. I think current QCs are, relatively speaking, at Eniac's level. And you know, Eniac to Galaxy S10 is quite a leap, so QCs might have some way to go still. I'm a bit sceptical of the field though and wonder if there truly is any difference between a QC and a probabilistic classical computer.mode_13h - Sunday, May 30, 2021 - link
Turing-equivalent just means that you can map the set of classical computation problems onto them, without regard for speed or space. It doesn't make them *good* at implementing classical computation, in a similar way that a lot of classical algorithms run poorly on something like a GPU.I think we'll probably still rely on conventional CPUs to implement conventional algorithms and just use quantum computers for large-scale, multi-parameter optimization problems. It doesn't mean they won't be revolutionary, just that they're not going to improve all of the things that faster conventional CPUs do.
GeoffreyA - Sunday, May 30, 2021 - link
Perhaps in the far future, once they've been shrunk enough and actually work, one could have a CPU that carries a little quantum unit. It could help with the tasks QCs excel at.mode_13h - Monday, May 31, 2021 - link
It's an idea, but entanglement seems far too fragile to do without enclosures on the caliber used for experiments on particle physics. Maybe I'll be proven wrong, but I expect the only way that 99.999% of the population ever encounters quantum computers is via the cloud.Anyway, we still need to solve the scaling problems, so they actually provide real benefits, before it's even worth focusing on ways to make them smaller, cheaper and remotely practical for an average person to have in their home.
GeoffreyA - Monday, May 31, 2021 - link
You're right. Not worth it as it stands. I think a probabilistic classical computer could be a more feasible approach. How close it would approach a QC, I don't know, but something like that could be up and running in no time. (If you can't get real-time lighting, go for lightmaps.) As for QCs, it's almost as if Nature is reluctant to co-operate. Some hint?I think a deeper insight, coming from theory, is also lacking and might lead to a better understanding of what exactly is going on at the lowest level. Clearly, superposition, entanglement, etc. are springing from some truth that's been caught in paradoxical classical terms. The secret lies with entanglement, I'd say; and concerning the other quantum mysteries, how much of them are spectres of the mathematics or physical reality.
GeoffreyA - Monday, May 31, 2021 - link
Anyhow, there are many itching to get their hands on quantum computers because of the implications for security, and they'll chase it to the ends of the earth. Emperor's New Clothes? Only time will tell.mode_13h - Monday, May 31, 2021 - link
> there are many itching to get their hands on quantum computers because of the implications for securityNot only that, but there are lots of high-value optimization problems. I see references to things like material science and drug discovery. D-Wave is also talking about deep learning, although that would have to be a very small model to train on their current-gen systems, I think.
GeoffreyA - Thursday, June 3, 2021 - link
Lots of biological systems, even plants, make use of quantum features, so these sort of computers could be a better fit for simulating them. I expect it would aid physics too, especially with simulating the quantum side of black holes.mode_13h - Monday, May 31, 2021 - link
> As for QCs, it's almost as if Nature is reluctant to co-operate.I don't know, but I think it'd be worthwhile if it gets the private sector to actually contribute to basic research, for once.
mode_13h - Sunday, May 30, 2021 - link
Speaking of Turing Machines, here's a paper on Neural Turing Machines, by some folks at Tenstorrent: https://arxiv.org/abs/1410.5401GeoffreyA - Sunday, May 30, 2021 - link
Much obliged for the links. I will look at each of them. In return, here's something I found yesterday:https://www.quantamagazine.org/artificial-neural-n...
Gomez Addams - Tuesday, June 1, 2021 - link
Except that's not a tough question. For these two guys it's about the most irrelevant question ever. They are not in the consumer-desktop processor business so how in the wide world of sports does that have anything to do with them?mode_13h - Wednesday, June 2, 2021 - link
Yeah, I thought about making that point. I guess it's tangentially relevant for Jim, since he just left Intel about 1 year ago, presumably working on consumer-desktop processors (among others). But, it really had nothing to do with the focus of the article.Anyway, the thread had already taken off, by then. I figured I'd play along.
movax2 - Thursday, May 27, 2021 - link
https://images.anandtech.com/doci/16709/ScaledML.p...Jim is air-playing the bass guitar here!!!
Slaping it hard!
mode_13h - Friday, May 28, 2021 - link
LOL, good call!webdoctors - Thursday, May 27, 2021 - link
Great article, cool to hear about a company starting in the a literal basement of Toronto (shows you how crazy their real estate prices are he still can't get out of his mom's basement). Kinda sad to hear about the internal management of AMD and Intel being like this (it sounds like they're hiring suits rather than real engineers for upper management which is a big red flag):AMD:
I wasn’t a Senior Fellow, I was actually a director - everybody at the time kept saying that nobody understands why I’m only a director and why am I not a fellow, or a senior fellow. That was a common theme there, but I guess I fit in more with these technical folks and they said there are a lot of organizational challenges to getting anything serious done. I thought that it was better [for me to be] positioned somewhere you have a bit of reach into both.
Intel:
Intel, people were surprised a Senior VP was grilling people on how they manage their batch queue, or what the PDK is, or what the qualification steps were, or how long the CAD flows took to run, or what the dense utilization was in the CPU, or how the register file was built. I know details about everything, and you know I care about them, I actually really care. I want them to look good. A friend of mine said, if it doesn't look good, it isn’t good. So when you look at the layout of a processor, or how something is built, it's got to be great. You can't make it great if you don't know what you're doing. It’s bad if somebody gives you like a five-level abstraction PowerPoint slide that said, things are going good, and we're improving things by 3% per year - I've been in meetings where somebody said 5% better, and I said ‘better than what? Better than last year, or better than the best possible, or better than your first-grade project?’. They literally didn't know what - they'd been putting 5% better on the slide for so long, they forgot what it was.
TERRIBLE.
Fataliity - Thursday, May 27, 2021 - link
That's why AMD back then had... Bulldozer, and Intel ended up with 14nm for like 7 years.The people up above them were marketing oriented, not engineering. Easy to BS someone that doesn't understand what your saying to them.
lmcd - Thursday, May 27, 2021 - link
While obviously Bulldozer is the highlight "problem," worth remembering that AMD spent 3 generations on 32nm and a 4th on 28nm bulk. Bulldozer wasn't just one bad design cycle.WaltC - Friday, May 28, 2021 - link
Guy, The President of AMD is an MIT engineer...;) AMD is filthy with engineers. It wasn't suits who put the company where it is today and it won't be suits that will keep them there.mode_13h - Friday, May 28, 2021 - link
> it sounds like they're hiring suits rather than real engineers for upper management which is a big red flagSo, um, maybe you didn't hear about this, but both Intel and AMD are now run by engineers.
As for your quote of Ljubisa's, that seemed pre-Zen. Obviously, AMD had issues at that time.
Jim didn't say the second part of your quote was specifically at Intel. If anything it sounded more like the beginning of his second stint and AMD, but he could've just been taking those examples from other presentations he sat through, in his long career.
GeoffreyA - Thursday, May 27, 2021 - link
Fantastic interview! Only halfway through but was grateful for the remarks on Zen's development; and I see there's a whole load of nutritional goodness up ahead. RISC-V, the works. Can't wait to read the rest (but eyes are sadly closing this evening).mode_13h - Friday, May 28, 2021 - link
The light it shed on the development of Zen and basically AMD's whole transformation was nothing short of amazing! This interview will quite likely be cited by tech historians chronicling that era.I think it's a fine line to walk, when trying to push people to step up and do better work. You mostly have to challenge, while being supportive. And you don't want to ruffle too many feathers and make enemies, or compromise your vision for the sake of consensus. But, you also don't want to seem impossibly demanding, so you need to bring people along towards your vision and not just cram it down their throats.
I think that's why having a team of experts who can & have all walked the walk is so powerful. Because, you can really challenge anyone who tries to throw up roadblocks, make excuses, or dig in their heels. And in tearing down opposition, you can also impress on everyone else what can be accomplished by a relentless and creative pursuit of solutions (instead of excuses).
*sigh*. It makes me wish I could work on a team like that. Sadly, the only inspiring leaders I've worked under were borderline jerks. And their plans tended to be half-baked, at best.
GeoffreyA - Sunday, May 30, 2021 - link
"The light it shed on the development of Zen"Yes, that was the golden part for me. The drawing back of the curtains, if only for a moment. From those few paragraphs, I can imagine the spirit of despair that had taken hold of AMD, along with their lack of belief. I'm glad Jim came in and shook them all up. And the right sort of man to put those geniuses in their places who were trying to block Zen. I think it's fair to say they'd have been content carrying on with Bulldozer.
GeoffreyA - Sunday, May 30, 2021 - link
"Sadly, the only inspiring leaders I've worked under were borderline jerks."Their personalities tend to be touched with an autocratic, self-centered ruthlessness, which, unfortunately, can lead to success but hurts many along the way. (I've read Steve Jobs was like that.) Funny thing is, folk fall at such characters' feet. Political/military leaders like this are famous.
mode_13h - Sunday, May 30, 2021 - link
In my case, they had a vision and a passion that would inspire others, but tended to be dismissive of people who didn't agree and could be rather petty an undiplomatic. I think that's kind of the norm, for the tech industry.It takes some patience and energy to bring people along with you, but sometimes they ask good questions or have good ideas, in the process. Plus, you can get more people on board and pulling in the same direction than with a sharp-elbowed approach. So, I think the key is to see that someone is digging in their heels and then to find a good way around them without too much fuss.
People seem to think a good leader has to be like Steve Jobs and break a few eggs to make a good omelet. I think a more avuncular approach is probably better, though it shouldn't be confused with being obsequious. You want to challenge people, in a supportive manner, and give them a real chance to come along -- yet, without compromising in your goals.
GeoffreyA - Monday, May 31, 2021 - link
Wholly agree. For my part, when a leader/manager treats you as a human, as an individual, tries to listen to what you've got to say, and considers it, whether he or she agrees or not---that's a person one desires to serve, follow, and please. I suppose it all comes down to their manner, even the tone of voice, and their being humble and down-to-earth. It comes naturally to some. Others haven't got it. A true leader does take charge and lead (without compromise on certain points), but doesn't put down others. Treats them with dignity, rather. Such a leader, one loves.KAlmquist - Saturday, May 29, 2021 - link
This was much better than the interview with Keller when Keller was at Intel. My guess is that for the Intel interview Keller had instructions from Intel's PR department on what to say and what not to say. So he was quite limited in what he could say about what he was doing at Intel, and he was no doubt prohibited from saying anything whatsoever about his work at AMD. It's also possible that he didn't like working at Intel very much.In this interview, he's talking pretty freely about stuff he's passionate about.
GeoffreyA - Sunday, May 30, 2021 - link
Yes, I vaguely remember that interview. It didn't have much at all. O, Intel.GeoffreyA - Sunday, May 30, 2021 - link
From the present interview, I get the impression that, at Intel, there's a lack of proper communication between the different levels (along with too big teams). Perhaps that's partly why they're in the state they are today. Hopefully, Mr. Gelsinger can stir things up there.blppt - Thursday, May 27, 2021 - link
I don't think its hyperbole to say that Jim Keller saved AMD.lmcd - Thursday, May 27, 2021 - link
Excavator vs Zen 1 is almost as big as Nehalem to Sandy Bridge.blppt - Friday, May 28, 2021 - link
Whats astonishing is that Intel was so screwed up he couldn't get anything done when there.ABR - Friday, May 28, 2021 - link
AMD was small enough that it _could_ be saved. Intel? Forget it.Anyway, what a great interview! It brings real insight into what makes Jim Keller who he is.
frbeckenbauer - Friday, May 28, 2021 - link
You have to look at the timeline, Jim Keller worked on products that haven't been released even announced yet.blppt - Friday, May 28, 2021 - link
All the rumors i've heard was that he essentially left Intel because they wouldn't do the radical changes he wanted, so my guess is that anything Intel releases that was designed by him would be underwhelming.mode_13h - Friday, May 28, 2021 - link
> All the rumors i've heardIn the comments thread soliciting questions for the next interview segment, Ian confirmed that Jim left for family reasons (and yes, *real* family reasons, even though that's the usual line people give). Unfortunately, it got deleted after the interview happened.
However, if you read the article, he had put his own money into Tenstorrent, before he even joined Intel. And after hearing about how many doors were opened to them after he started working at Tenstorrent, I think he really didn't need any other reason to go there than to help them succeed.
Plus, he's notorious for leaving places after only a few years. His stint at Intel was short, but not totally out of character for him.
mode_13h - Friday, May 28, 2021 - link
> Intel was so screwed up he couldn't get anything done when there.And how do you know that?
I'm sure it'll be at least a couple more years before the stuff he was working on sees the light of day.
mode_13h - Friday, May 28, 2021 - link
> Excavator vs Zen 1 is almost as big as Nehalem to Sandy Bridge.WTF? Nehalem wasn't bad at all. I think you mean more like Pentium 4 to Core 2.
lmcd - Friday, May 28, 2021 - link
Excavator wasn't awful either lol, it had way better power characteristics than any prior Bulldozer derivative and did quite well when it first came out, but its biggest benefits were platform changes for AM4 that it didn't utilize all that well, but its successors did.This is also a good description of Nehalem vs Sandy Bridge. Nehalem's biggest changes were platform-level in migrating toward an SoC-like design. Its performance benefits were nice over the previous generation but the big overhaul of Sandy Bridge is what people remember for a reason -- they were more important.
The relative performance positions of AMD vs Intel in those comparisons is the primary reason you're objecting imo. Doesn't change the internal jump, which is my point. The wild ride of micro-op caches cannot be understated. Neither can the jump from the weird "permanent SMT" Bulldozer module to Zen's CCX.
mode_13h - Saturday, May 29, 2021 - link
Okay, that was a repectable rebuttal. Thanks for confirming your original statement.Wereweeb - Thursday, May 27, 2021 - link
Thanks for sneaking that silicon photonics question in. I suppose he's right, tech "enthusiasts" are always talking about the next big thing that will revolutionize everything while the foundries are busy actually improving what already works, and people like Jim are re-thinking how to use the different tools to actually do what they need being done.Still, while I find his enthusiasm for technology great, I still can't but be skeptical of AI. At some point they'll be able to tell which kind of books you might be reading just by the words you use or the way you structure your sentences. Imagine a totalitarian dictatorship having that kind of power. Perhaps is the kind of power that humans don't deserve to wield.
mode_13h - Saturday, May 29, 2021 - link
Hopefully, the second part of the interview will go into the ethics and other concerns around AI. As we can appreciate that Jim is a thoughtful guy, I'd expect he's contemplated it in some depth.GeoffreyA - Saturday, May 29, 2021 - link
Oh yes. AI is going to be much abused in time to come. And it's happening in a milder way at present.GeoffreyA - Saturday, May 29, 2021 - link
Can't resist adding that we can expect Google, Facebook, and Co. to use AI very ethically. Why? Because they care about us.Tomatotech - Saturday, June 5, 2021 - link
@wereweeb> At some point they'll be able to tell which kind of books you might be reading just by the words you use or the way you structure your sentences.
They already can. Check out OpenAI’s GPT-2. You can feed it a text snippet and it will write an article using the same phrasing and topic style. I believe a similar or later version had to be withdrawn from general release because it was too realistic and made it too easy to mass-produce political propaganda.
https://app.inferkit.com/demo
Remember this demo is from TWO years ago which is an eternity in AI terms.
plopke - Thursday, May 27, 2021 - link
lot of things were out of my league but still a fun read :)flgt - Thursday, May 27, 2021 - link
Thank you for the interview. I’ve really come to appreciate what makes Jim Keller so great after reading/watching some of his interviews. Working in the tech industry for awhile now and working for what has become a mega-corporation, I can relate to the things he says and some of his comments about failing cultures. It really does go beyond the technology. It’s kind of scary listening to the AMD story and knowing that most corporations will ride failing leadership/culture to the bottom until there is enough desperation to change things. And most times you don’t have a JK to save you. I wish I had the chops to be part of his circles.blppt - Friday, May 28, 2021 - link
Even if all you know about him is that he's the primary reason x86-64 and Zen exist, that seals his fate as a HoFer.Timoo - Sunday, May 30, 2021 - link
?Hofer: One who cannot do anything right, or does it in such a way that he may be considered to be mildly retarded.
Are you sure you mean that? (this is from Urban Dict)
blppt - Sunday, May 30, 2021 - link
Thats why I capitalized the 2nd F---Hall of Famer. ;)grrrgrrr - Thursday, May 27, 2021 - link
It's really nice for tenstorrent to make their accelerator cards available to the public. Availability of hardware is crucial to reproducibility of AI research. Looking forward to getting one!mode_13h - Saturday, May 29, 2021 - link
If he said exactly that, I must've missed it. However, their website says they are already bringing a devcloud online, and it currently supports inferencing workloads. They also have a link to apply for early access to Greyskull.https://tenstorrent.com
mode_13h - Saturday, May 29, 2021 - link
Oh, yeah... their FAQ definitely confirms they're selling Greyskull PCIe cards to the public.https://tenstorrent.com/faq/
mode_13h - Saturday, May 29, 2021 - link
List prices and hardware specs are here:https://tenstorrent.com/grayskull/
Timoo - Sunday, May 30, 2021 - link
That's actually quite cool!I mean, according to their investment rounds, they are burning through the millions. And still [coming soon] you can simply buy a card for 1000$, download the software for free and have a go with it! That's só cool!
fliguy84 - Friday, May 28, 2021 - link
I don't really get the flight booking situation that LB mentioned. Did he accidentally book a flight one week earlier than on a date he was supposed to travel?mode_13h - Friday, May 28, 2021 - link
Yeah, I wish that had been a little clearer, on a couple points. I'm inferring that Jim's advice to him was to buy a ticket on his own dime, because missing those meetings would do too much damage to his standing in the organization. But, it'd be nice if it clarified exactly what Jim was referring to and what Ljubisa ultimately did about it.That said, it was a tangent, and its purpose was really just to show how Jim would give him career advice. So, we needn't worry too much about the particulars.
Exotica - Friday, May 28, 2021 - link
Fascinating. Keller comes across a a breath of fresh air, like someone who wants to change the (silicon) world. Seems like he has a ton of good energy. From how Keller describes intel and the bureaucracy, it sounds like it was killing him slowly inside. Hopefully he left a mark and we see good things from the fruit of his labor, similar to what we saw at AMD.WaltC - Friday, May 28, 2021 - link
I think Keller's fine--like the guy just fine. But at some point people need to get past the "hero" syndrome, imo...;) Playing musical chairs with your employment is fine when you have the necessary funds under your belt to ensure that you don't starve no matter what happens. Most company employees don't have that privilege. But getting to where AMD is today is 100% a collaborative effort undertaken by thousands of employees and engineers all contributing something of importance. Keller's now with another company so naturally he will promote himself and his new company while he criticizes the other guys...it's just business 101. What I really, really like about AMD is that the company isn't held together by "rockstar" engineers and programmers--indeed, teamwork happens to be at a premium at AMD, imo. Intel, of course, is pretty much a lost cause--the company is so waterlogged and rotten with turf wars and monopoly thinking it's a miracle they can produce much of anything competitive. The company had to bring an old man out of retirement because nobody there had an idea of what to do next regardless of how much money they were drawing down. Really pitiful, imo.flgt - Friday, May 28, 2021 - link
I don’t think JK would say AMD’s success was due to one person. He’s always come off as an anti-prima donna/blue collar kind of guy to me. But the reality is that for every really good engineer with great ideas, there are 2 or 3 really bad engineers. And the bad ones would rather fall on their sword than ever admit they are wrong. If you don’t have guys like Keller to 1) correctly sort good and bad ideas 2) find and enable good engineers 3) check/block/remove bad engineers by any means necessary 4) set the direction and expectations for execution, your company will under perform. Sounds easy but it’s amazing how hard it is to find those X-factor guys. So why should the average employee care? Because in the commercial world it makes the difference between growth/bonuses and you losing your job. If you’re working in academia or defense, maybe you don’t care. But a lot of people appreciate a track record of delivering good products.mode_13h - Friday, May 28, 2021 - link
> Our initial target isn't to get some huge contract, it is to get 100 programmers> using our hardware, programming it, and living with it every day.
Sadly, AMD still doesn't get this. Nvidia got it, and it helped give them their lead in machine learning. But AMD's Linux driver stack for GPU-compute has been absent or broken for RDNA cards, since day one.
All AMD's GPU-Compute division seems to care about is HPC and hyperscalers. They don't seem to understand that it's kids in university who are pioneering the new frameworks and algorithms. And they're using gaming cards, not $10k compute accelerators!
AMD's first priority should've been top-to-bottom hardware support, for their entire software stack! Even if it meant missing out on a couple supercomputer wins, the long-term payoff would outweigh that by orders of magnitude.
GeoffreyA - Saturday, May 29, 2021 - link
Not too long ago, a basic thing like hardware encoding was broken on AMD GPUs in Handbrake. I remember the HB team contacted AMD, and it was fixed after that, requiring updates to both the driver and HB.Tomatotech - Saturday, June 5, 2021 - link
I got from this that they don’t yet have more than 15 people in the world able to program their hardware. Still very early days, when counting everyone who knows how code for your new stack just means taking off your socks.*he said his team has 15 people on it.
mode_13h - Sunday, June 6, 2021 - link
> they don’t yet have more than 15 people in the world able to program their hardware.I think his point was that he was drawing a contrast with other AI startups, who just want to deliver their solution as a black box that customers only use via supported deep learning frameworks.
Tenstorrent wants people to open the box and tinker inside. Based on their website, they seem interested for people to use their chips even for non-deep learning applications.
mode_13h - Friday, May 28, 2021 - link
> LB: You get the flexibility, definitely. Adding programmability isn’t free as usual,> so there’s a bit of sacrifice there.
This is why I've long thought GPUs had an advantage over a lot of the AI accelerators. GPUs give up a bit of performance, but they offer a lot in programmability. And they're very mature, in that respect. There's obviously code that runs better and code that runs worse on them, but they're still quite versatile, compared with a lot of what's out there, and you can usually find a way to still get decent performance.
frbeckenbauer - Friday, May 28, 2021 - link
the move to seperate compute and gaming architectures means they have very little reason on porting their new software stack to old gaming cards, I'm afraid.mode_13h - Friday, May 28, 2021 - link
You're replying to my above point about RDNA? If AMD thinks they don't care about RDNA as a viable platform for compute, then they're basically saying they want AMD GPUs to be about as relevant to the future of computing as the other HPC-exclusive players you've probably never heard of.Nvidia gives away full CUDA support, on their *entire* hardware line, from $100 Jetson boards on up. Sure, some of their GPUs have stuff a bit nerfed, like reduced NVENC streams and fp16 tensors running at half speed, but it's still usually fast enough to be worthwhile, and certainly you can use it as a software development platform.
iranterres - Friday, May 28, 2021 - link
LOL at the comic sans ms at the first picturemode_13h - Friday, May 28, 2021 - link
JK> We are trying to meet the software guys where they are at, because theyJK> just want to write code. They also want to understand the hardware
JK> so they can drive it, but they don't want to be tortured by it.
So true. You want hardware that's interesting enough to be exotic and to let you push the envelope, but not so byzantine that you spend 98% of your time and energy catering to it, rather than focusing on your software and algorithms.
mode_13h - Friday, May 28, 2021 - link
JK> I know details about everything, and you know I care about them, I actually really care.I think that's the secret. That's how you earn the respect of the rank-and-file. You show that you understand and care about their work, while also demonstrating competence they can appreciate. It also helps keep them on their toes, letting them know that they're not going to get by with second-rate work.
JK> A friend of mine said, if it doesn't look good, it isn’t good.
JK> You can't make it great if you don't know what you're doing.
Exactly. There's no room for BS, when you're reaching for excellence. You not only need to understand how something works, but how and why it's better than all the alternatives and why it can't be even better.
JK> I've been told at various points in my career to focus on that high-level picture of
JK> managing, but I always like to get into the details
I've worked for managers that spent all their time in the details and didn't do enough managing. That was not good. At some point, the opportunity cost of managers not managing becomes greater than the value of the technical contributions they can make. In Jim's case, it sounds more like he was acting as a consultant, rather than trying to do the actual work of an individual contributor.
But, I've also seen middle-managers try to lead an engineering organization, without the core technical chops. They didn't know when they were being fed a load of BS, didn't hire the right people under them, and tended to make decisions more for political than technical or sound business reasons. And that's the kind of stuff that can rot a department from the inside out.
mode_13h - Friday, May 28, 2021 - link
JK> abstraction stack - there's the high level, there's the medium, and there's a low level.JK> Yes, you do need to know a lot about all of them,
I'd say: pick your level and own it. Master it as well as anyone. But, you've got to know enough about the other levels that you at least know when to go there, and can pick up what else you need, as necessary.
I've known a lot of generalists, and I've got to say that they're rarely as interesting or impressive as the specialists. Just don't be afraid to go outside your domain.
Timoo - Sunday, May 30, 2021 - link
They are 2 different kinds of workers.As a generalist myself, I try to stay up-to-date with all the fancy stuff the specialists are working on. I use this knowledge to integrate it with the work of other specialists. While the specialists use it to advance in their field. Of which they then again keep me up to date, so I can integrate it with the work of others. And so on.
2 different ways of working.
Timoo - Sunday, May 30, 2021 - link
ps. For a specialist, there are só many opportunities to advance, só many options. But there's only a limited amount of time, so they have to choose their next steps carefully. This is where people like me try to be of value; by having an overview, I can make them realise what next step could be beneficial to the company as a whole or to the platform they are working in, and which would be wáy out of line and/or very hard to integrate.Mais bon. C'est la vie.
mode_13h - Monday, May 31, 2021 - link
Thanks for sharing your experience. I'm aware that some places with a lot of specialists tend to pair them with generalists, who can help integrate their research into real products.It's rare, but you sometimes see a specialist with all the versatility of a generalist.
mode_13h - Friday, May 28, 2021 - link
JK> I heard from many people that Jack could get to the bottom of anything.A lot of it just comes down to fearlessness and not being afraid to ask a dumb question or seem ignorant. And if you ask enough questions, you start to learn which are the key ones. That and persistence will get you most of the way there.
mode_13h - Sunday, May 30, 2021 - link
One question I have is why keep the ARC cores, once they start integrating RISC-V. I guess that's probably not yet set in stone (or lithography masks, as it were), but I'd expect them to just go RISC-V everywhere they don't have a good reason not to.mode_13h - Sunday, May 30, 2021 - link
BTW, I'm guessing the ARC cores are these: https://en.wikipedia.org/wiki/ARC_(processor)It's interesting to note their lineage traces back to the FX processors in a couple Super NES game cartridges. I'm guessing they ended up in Grayskull via its FPGA-based predecessor, as they now seem to be part of Synopsys' licensable IP library.
mode_13h - Sunday, May 30, 2021 - link
And, from a user's perspective, I'm wondering a lot about how DRAM fits into the programming model. Presumably, it can be accessed from any node in the graph, even if doing so isn't exactly fast.I think DRAM is, among other things, key to their scaling story. It's a fundamental part of how they could take a processing graph and scale it down to a single chip, even if the entire set of weights or some intermediate state is too big to fit in the chip's SRAM. In that sense, how it's accessed could be an easily-overlooked means of differentiating themselves vs. others taking a similar sort of dataflow-oriented approach.
mode_13h - Sunday, May 30, 2021 - link
And speaking of their competition, I'm struck by some apparent similarities between their approach and that of Cerebras and perhaps even Habana and Graphcore. Basically, the idea of building a dataflow processor and compiling a processing graph to fit either one chip or span multiples, via high-bandwidth networking and a roughly mesh-like topology.To the extent that description is accurate, it could be something akin to the formula the industry has settled on for building fast GPUs (i.e. wide SIMD + deep SMT), or even fast CPU cores.
Igor_Kavinski - Tuesday, June 1, 2021 - link
Amazing. So much info jampacked into this interview. And the anecdotes! Can't wait for the one on one interview with Jim Keller! You go, Ian!mode_13h - Wednesday, June 2, 2021 - link
+1Tomatotech - Saturday, June 5, 2021 - link
Excellent interview - I loved it!I put off reading it for a week as it was so long but I’m glad it read it in the end. Not a word was wasted. Some other interviews on AnandTech have been full of corporate speak and waffle so I was reluctant to read this one. But this one really delivers. Excellent! I suspect you’ll get a lot of hits coming back to this article over the next few years.
Looking forward to the next one with Jim!
PS I lol’d at the question: ‘what’s the limitations of your new chip?’ Just like being asked at a job interview ‘What’s your greatest weakness?’ I’ve sometimes asked that at interviews for kicks & to see how they deal with difficult questions - and how well prepared they are. The answer itself doesn’t really matter as long as it’s delivered well. Usually get something like ‘I work too hard’ or ‘I care too much about the job’. I don’t ask it any more as we sometimes get (excellent) disabled candidates and it would be wrong (and potentially discriminatory) to ask them ‘what’s your greatest weakness?’. There are other ways of asking difficult questions in interviews.
mode_13h - Sunday, June 6, 2021 - link
> I don’t ask it any more as we sometimes get (excellent) disabled candidatesYou could use a simple variation: "What's one big decision you've made in a project that you'd now make differently?"
In some ways, it's better than your original question, since it reveals how thoughtful and invested in their projects they tend to be, as well as their sense of agency.
Sara adams - Saturday, June 5, 2021 - link
Jim's super cool and also very respected in the industry. He was a big inspiration for me early on in my career. I was a fairly junior CPU performance engineer at AMD when Jim joined AMD in 2012 (he'd worked at AMD earlier in his career too). I worked on the Zen program, which was being led by Jim. I saw him at my gym one morning, deadlifting 275 lbs (or maybe more) and I went up to him and introduced myself. He was super friendly and continued to be so whenever I'd run into him. We'd exchange our personal bests in lifting. IIRC, he was in his mid 50's then. Given the similarity in our backgrounds (CPU design) and his professional achievements and his amazing discipline, it was a no-brainer for me to aspire to be like him. www.ilmibook.com says This is a man at the cutting edge of his field.mode_13h - Sunday, June 6, 2021 - link
Neat anecdote, but I'm for sure not visiting that website.lsdigital - Saturday, August 7, 2021 - link
Thank you for informative article .are you searching for help to get ideas about the Digital marketing service? If yes, then you should try the..<a href="https://lsgodigital.co.uk/5-ways-to-digitally-prom... Ways To Digitally Promote Your Business </a>lsdigital - Saturday, August 7, 2021 - link
Thank you for informative interview .are you searching for help to get ideas about the Digital marketing service? If yes, then you should try the.<a href="https://lsgodigital.co.uk/5-ways-to-digitally-prom... Ways To Digitally Promote Your Business </a>