I was todays years old, when I found out that Lisa Su (CEO of AMD) and Jensen Huang (co-founder and CEO of NVIDIA) are relatives!
If you can't do a merge, it's good to have family onboard ;-)
They didn't know about their relation until much later, but if they had Jensen would have been the "cousin you don't want to be like" - he went to Oregon State and worked at a Denny's while Lisa Su went to Bronx Science and on to MIT.
> he went to Oregon State and worked at a Denny's while Lisa Su went to Bronx Science and on to MIT.
Seems like a bad vibe to imply someone shouldn't aspire to go to state school or work a humble job for money to get through it, even though given both options, indeed they may dream about the fancy one. Denny's has the best milkshakes anyway and state school is probably a much more sensible place to attend.
Although he did graduate high school two years early, so he had the intuition. Maybe his parents thought working food service for a bit was a rite of passage.
(-; Indeed always warming to the core to see productive emulation between parallel lines! I'm sure after all their achievements, neither of them wastes time pondering woulda, shoulda, cuda…
Yep, I think that's the best quote of the interview and the runner up is the Napster quote.
People seems to have forgotten that Intel used to had StrongARM in their lineup, with the same logic Intel is a computing company not x86 company, similar to AMD [1].
For one of the latest trends in the computing landscape please check this new Embedded+ platform recently introduced by AMD [2]. I'm biased towards IoT but I've a very strong feeling that IoT is the killer application of AI, similar to AI is a killer application of HPC as the latter statement was mentioned by Lisa Su during the interview. The turning point for IoT however is yet to happen, it will probably happened when the number of machine nodes talking to directly to each other, surpassing the number of machine nodes talking to human, then it will be its ChatGPT moment.
It has huge implications because it makes competing in that market much harder due to licensing issues. Intel and AMD have a duopoly on the x86 market which compromises a huge chunk of server and personal computing, but that is changing fast. If they go ARM (or risc-v or whatever) they will have more competition to contend with, including their existing cloud computing clients designing their own chips and fabbing them with other foundries.
It is good to mention the AMD's Steam Deck CPU [1] running in the Steam Deck [2] and not less important that the Steam Deck also has Linux (and KDE) incorporated [3].
I have endless admiration for Lisa Su, but lets be honest, the reason AMD and Nvidia are so big today is that Intel has had amazingly bad management since about 2003.
They massively botched the 32-bit to 64-bit transition....I did some work for the first Itanium, and everybody knew even before it came out that it would never show a profit. And they were contractually obligated to HP to not make 64-bit versions of the x86....so we just had to sit there and watch while AMD beat us to 1 Gigahertz, and had the 64-bit x86 market to itself....
When they fired Pat Gelsinger, their doom was sealed. Thank God they hired him back, but now they are in the same position AMD and Nvidia used to be in: Intel just has to wait for Nvidia and AMD to have bad management for two straight decades....
You're talking about very old events. Intel made mistakes during 2000-2005 (which allowed the rise of Opteron) but they crushed it from 2006-2016. Then they had different problems from 2016-2024.
Intel went into Rest & Vest mode after Haswell (4000 series) in 2013 - they hardly improved until 2020 - 6 lost generations of relabeled CPUs like 13000 - 14000 series (lots of money paid for innovative new igpu marketing names like 520 and 620!). They found out that community college talent does not make good employees in the C-Suite ... The iris pro 5200 (2013) wasn't improved upon until 2020 !
I don't mean to take away from Intel's underwhelming management.
But regardless, Keller's Athlon 64 or Zen are great competitors.
Likewise, CUDA is Nvidia's massive achievement. The growth strategy of that product (involving lots of free engineer hours given to clients on-site) deserves credit.
// I don't mean to take away from Intel's underwhelming management
chuckle lets give full credit where credit is due :-)
Athlon was an epochal chip. Here's the thing though---if you are a market leader, one who was as dominant as Intel was, it doesn't matter what the competition does, you have the power to keep dominating them by doing something even more epochal.
That's why it can be so frustrating working for a #2 or #3 company....you are still expected to deliver epochal results like clockwork. But even if you do, your success is completely out of your hands. Bringing out epochal products doesn't get you ahead, it just lets you stay in the game. Kind of like the Red Queen in Alice in Wonderland. You have to run as fast as you can just to stay still.
All you can do is try to stay in the game long enough until the #1 company makes a mistake. If #1 is dominate enough, they can make all kinds of mistakes and still stay on top, just by sheer market inertia. Intel was so dominate that it took DECADES of back-to-back mistakes to lose its dominate position.
Intel flubbed the 32-64 bit transition. On the low end, it flubbed the desktop to mobile transition. On the high end, it flubbed the CPU-GPU transition.
Intel could have kept its dominate position if it had only flubbed one of them. But from 2002 to 2022, Intel flubbed everysingle transition in the market.
Its a measure of just how awesome Intel used to be that it took 20 years....but there's only so many of those that you can do back-to-back and still stay #1.
I began supporting AMD as my choice for a gaming CPU when they began trouncing Intel in terms of performance vs. total draw power with an attractive price point around 2016 or so.
Then, a wave of speculative execution vulnerabilities were discovered / disclosed, resulting in an even larger differential for performance and power use after the SPECTRE patches were applied.
Considering this, I'm not sure that it's fair to cast the successes of AMD as mere failures from Intel. Dr. Su is simply a brilliant engineer and business leader.
Well geez, Intel might have amazingly bad management but where on the scale do we put AMD not giving any fucks about challenging the CUDA monopoly for going on 10 years now?
Instead they put mediocre people on things like OpenCL which even university students forced to use the mess could tell was going nowhere.
AMD basically did not have any money from 2010 until 2020. They were teetering near death. No money for r&d no money for new features in their gpus no money no money no money. There excavator architecture was extremely foolish trying to support multiple instruction dispatches with only 1 ALU and FPU per core!
This was corrected with Ryzen and the last 4 years they have been able to pay off their massive debts, especially with the 5000 series!
Going forward I expect we will see much more innovation from them because now they have some spare cash to spend on real innovations hopefully in hardware AND software! Note that Intel is much worse than AMD in software!
This article inspired me to check out the respective market caps for Intel and AMD. Hell of a turnaround! I remember the raging wars between the Pentiums and the Athlons. Intel won. The GPU wars between Nvidia and ATI; Nvidia won. Thereafter AMD, the supposed loser to Intel, absorbed ATI, the supposed loser to Nvidia. But I love that the story didn't end there. Look at what AMD did with Sony PlayStation (extensively discussed in this interview)...and that's without getting into the contemporary GPU AI revolution that's likely driving AMD's $250+ billion market cap. Epic!
And yet they still can't solve the problem of their GPU driver/software stack for ML being much worse than NVidia's. It seems like the first step is easy: pay more for engineers. AMD pays engineers significantly less than NVidia, and it's presumably quite hard to build a competitive software stack while paying so much less. You get what you pay for.
Well this is glaringly obvious to whole world, and Nvidia managed to get it right. Surely a feat that can be repeated elsewhere when enough will is spread over some time. And it would make them grow massively, something no shareholder ever frowns upon.
Exactly. ROCm is only available for the top tier RX 7900 GPUs and you’re expected to run Linux.
AMD was (is) tinkering with a translation layer for CUDA, much like how WINE translates directX. Great idea but it’s been taking a while in this fast paced market.
> AMD was (is) tinkering with a translation layer for CUDA
From what I understand, they dropped the contract with the engineer who was working on it.
Fortunately, as part of the contract, said engineer stipulated that the project would become open source, so now it is, and is still being maintained by that engineer.
Came here to say this. They only just recently got an AMD GPU on MLPerf thanks to a (different company), Tinycorp by George Hotz. I guess basic ML performance is too hard a problem.
I dunno, it a world where hardware companies like, sold hardware, and then software companies wrote software and sold that could be pretty nice. It is cool that Hotz is doing something other than contribute to an anticompetitive company’s moat.
AMD GPUs are pretty well represented on the TOP500 list (https://top500.org/lists/top500/list/2024/06/), which tends to feature computers used by major national-level labs for scientific research. AMD CPUs are even moreso represented.
* HPC tends to focus exclusively on FP64 computation, since rounding errors in that kind of use-case are a much bigger deal than in DL (see for example https://hal.science/hal-02486753/document). NVIDIA innovations like TensorFloat, mixed precision, custom silicon (e.g., the "transformer engine") are of limited interest to HPC customers. It's no surprise that AMD didn't pursue similar R&D, given who they were selling GPUs to.
* People tend to forget that less than a decade ago, AMD as a company had a few quarters of cash left before the company would've been bankrupt. When Lisa Su took over as CEO in 2014, AMD market share for all CPUs was 23.4% (even lower in the more lucrative datacenter market). This would bottom out at 17.8% in 2016 (https://www.trefis.com/data/companies/AMD,.INTC/no-login-req...).
AMD's "Zen moment" didn't arrive until March 2017. And it wasn't until Zen 2 (July 2019), that major datacenter customers began to adopt AMD CPUs again.
* In interviews with key AMD figures like Mark Papermaster and Forrest Norrod, they've mentioned how in the years leading up to the Zen release, all other R&D was slashed to the bone. You can see (https://www.statista.com/statistics/267873/amds-expenditure-...) that AMD R&D spending didn't surpass its previous peak (on a nominal dollar, not even inflation-adjusted, basis) until 2020.
There was barely enough money to fund the CPUs that would stop the company from going bankrupt, much less fund GPU hardware and software development.
* By the time AMD could afford to spend on GPU development, CUDA was the entrenched leader. CUDA was first released in 2003(!), ROCm not until 2016. AMD is playing from behind, and had to make various concessions. The ROCm API is designed around CUDA API verbs/nouns. AMD funded ZLUDA, intended to be a "translation layer" so that CUDA programs can run as a drop-in on ROCm.
* There's a chicken-and-egg problem here.
1) There's only one major cloud (Azure) that has ready access to AMD's datacenter-grade GPUs (the Instinct series).
2) I suspect a substantial portion of their datacenter revenue still comes from traditional HPC customers, who have no need for the ROCm stack.
3) The lack of a ROCm developer ecosystem means that development and bug fixes come much slower than they would for CUDA. For example, the mainline TensorFlow release was broken on ROCm for a while (you had to install the nightly release).
4) But, things are improving (slowly). ROCm 6 works substantially better than ROCm 5 did for me. PyTorch and TensorFlow benchmark suites will run.
Trust me, I share the frustration around the semi-broken state that ROCm is in for deep learning applications. As an owner of various NVIDIA GPUs (from consumer laptop/desktop cards to datacenter accelerators), in 90% of cases things just work on CUDA.
On ROCm, as of today it definitely doesn't "just work". I put together a guide for Framework laptop owners to get ROCm working on the AMD GPU that ships as an optional add-in (https://community.frame.work/t/installing-rocm-hiplib-on-ubu...). This took a lot of head banging, and the parsing of obscure blogs and Github issues.
TL;DR, if you consider where AMD GPUs were just a few years ago, things are much better now. But, it still takes too much effort for the average developer to get started on ROCm today.
Summary: AMD works if you spend 500m USD+ with them. Then they'll throw an army of their own software engineers into the contract who will hold your hand every step of the way, and remove all the jank for you. By contrast, since at least 10 years ago, I could buy any GTX card and CUDA worked out of the box, and that applied right down to a $99 Jetson Nano.
AMD's strategy looks a lot like IBM's mainframe strategy of the 80s. And that didn't go well.
Good summary. There was also the 2010's multivendor HSA and OpenCL software evolution directions that ended up losing other vendors on the way and many customers turned out to accept the proprietary Cuda.
I have a 7900 XTX. There's a known firmware crash issue with ComfyUI. It's been reported like a year ago. Every rocm patch release I check the notes, and every release it goes unfixed. That's not to go into the intense jank that is the rocm debian repo. If we need DL at work, I'll recommend Nvidia, no question.
Omg. I know this is mostly marketing speaking, but this is her reply when asked about AMD's reticence to software:
> Well, let me be clear, there’s no reticence at all. [...] I think we’ve always believed in the importance of the hardware-software linkage and really, the key thing about software is, we’re supposed to make it easy for customers to use all of the incredible capability that we’re putting in these chips, there is complete clarity on that.
I'm baffled how clueless these CEOs sometimes seem about their own product. Like, do you even realize that this the reason why Nvidia is mopping the floor with your stuff? Have you ever talked to a developer who had to work with your drivers and stack? If you don't start massively investing on that side, Nvidia will keep dominating despite their outrageous pricing. I really want AMD to succeed here, but with management like that I'm not surprised that they can't keep up. Props to the interviewer for not letting her off the hook on this one after she almost dodged it.
She could admit that they fell behind on this one and really need to focus on closing the gap now. But instead she says it's all business as usual, which assures me that I won't give their hardware another shot for quite a while.
She could say "We are in a compute gold rush and yet AMD's stock didn't gain anything in the last 6 months, so I hereby submit my resignation". That would work.
I'd love to know if any domain experts have a write up on what the the talent+ time+financial investment it would take for AMD to come up with with something that is a worthy rival to CUDA. Very curious to understand what the obstacles are.
~5 years. Medium-sized team in-house + hordes (hundreds, thousands) of engineers in the field helping clients on-site, writing code for them directly upstreamed to drivers, core libs, etc. (iteratively optimized in-house, ship feature, rinse and repeat). Story of the PlayStation SDKs, of DX too, but above all CUDA (they really outdid this strategy), now for cuDNN and so much more.
It takes incompressible time because you have to explore the whole space, cover most bases; and it takes an industry several years (about one "gen" / hardware cycle) to do that meaningfully.It helps when your platform is disruptive and customers move fast.
Maybe 3 years at best if you start on a new ideal platform designed for it from scratch. And can throw ungodly amount of money fast at it (think 5K low-level engineers roaming your installed base).
Maybe 10+ yrs (or never) if you're alone, poor, and Radeon (j/k but to mean it's non-trivial).
I’d say it mainly needs persistence and good execution (library support). NVIDIA has co-developed CUDA with their hardware, and largely stayed compatible with it, since around 2009, and around 2012 it first started taking off in the HPC space. Years later this enabled first their boom in crypto and then an even bigger one in AI. I don’t think this amount of R&D would be out of reach of today’s AMD (as NVIDIA wasn’t any bigger back then), but the backing of it needs to come from the very top.
First, they need to work with kernel devs to finally fix their drivers. Like, Nvidia used to be a "pain in the ass" here as well (that's a literal quote from Torvalds), so simply by contributing more than nothing, they could have taken the lead. But they definitely screwed this one up.
Second, they need to fix their userspace stack. ROCm being open source and all is great in principle, but simply dropping your source to the masses doesn't make it magically work. They need to stop letting it linger by either working with the open source community (huge time investment) or do it themselves (huge money investment).
This AI stuff has progressed a bit. Intel has been working on interesting stuff with OneAPI. It might be the case that things have progressed to the point where the primitives are well enough understood that you need something more like a good library rather than a good compiler.
In the end, more people seem to love BLAS than Fortran, after all.
I don't want a CUDA rival. I want to get the entire pile of CUDA code that is already written and run it on AMD GPUs without any kind of tweak or rewrite, and have it just work every time
Compatibility with existing code is very important. People can't afford to rewrite their stuff just to support AMD, and thus they don't
AMD is kind of trying to do this with rocm and HIP, but whatever they are doing it's not enough
My theory is that someone came up with the bright idea of allowing more open source in the stack and that that would allow them to get it all done via crowd sourcing and on the cheap. But if true it was a quite naive view of how it might work.
If instead they said let's take the money we should invested in internal development and build an open developer community that will leverage our hardware to build a world class software stack it might have been a little better.
> When I was working on TVM at Qualcomm to port it to Hexagon a few years ago we had 12 developers working on it and it was still a multiyear long and difficult process.
> This is also ignoring the other 20 or so developers we had working on Hexagon for LLVM, which did all of the actual hardware enablement; we just had to generate good LLVM IR. You have conveniently left out all of the LLVM support that this all requires as AMD also uses LLVM to support their GPU architectures.
> Funny enough, about a half dozen of my ex coworkers left Qualcomm to go do ML compilers at AMD and they're all really good at it; way better than I am, and they haven't magically fixed every issue
> It's more like "hire 100 additional developers to work on the ROCM stack for a few years"
This last statement sounds about right. Note that ROCm has over 250 repos on Github, a lot of them pretty active: https://github.com/orgs/ROCm/repositories?type=all - I'm sure an enterprising analyst who was really interested could look at the projects active over the past year and find unique committers. I'd guess it's in the hundreds already.
I think if you click through the ROCm docs https://rocm.docs.amd.com/en/latest/ (and maybe compare to the CUDA docs https://docs.nvidia.com/cuda/ ) you might get a good idea of the differences. ROCm has made huge strides over the past year, but to me, the biggest fundamental problem is still that CUDA basically runs OOTB on every GPU that Nvidia makes (with impressive backwards and in some cases even forwards compatibility to boot https://docs.nvidia.com/deploy/cuda-compatibility/ ) on both Linux and Windows, and... ROCm simply doesn't.
I think the AMD's NPUs complicate things a bit as well. It looks like it's its currently running on its own ONNX/Vitis (Xilinx) stack https://github.com/amd/RyzenAI-SW , and really it should either get folded into ROCm (or a new SYCL/oneAPI-ish layer needs to be adopted to cover everything).
But when Developers still continue buying NVIDIA for CUDA, because developers only target CUDA for their applications it is a chicken-egg scenario, similar to Linux vs Windows.
> Like, do you even realize that this the reason why Nvidia is mopping the floor with your stuff?
Are they?
Generally, my experience has been that AMD products generally just work even if they're a bit buggy sometimes, while Nvidia struggles to get a video signal at all.
Seems perfectly reasonable to focus on what matters, while Nvidia is distracted by the AI fad.
I'm not talking about gaming, I'm talking about general purpose computing (although even for gaming your statement is pretty bold). Since she's CEO of a publicly traded company, it seems pretty weird that she would ignore the fields where the money is at, while Nvidia becomes the most valuable company in the world. So she's not just ignoring developers' wants but also her stockholders'.
even if the LLM thing is an "AI fad" there are many other things that ML is used for that matter (to the people spending real money on GPUs - think A6000, H100, not gamer cards)
She worked as a researcher in the field for decades. Moore's law only kept up because of some of the techniques she developed.
> During her time at IBM,[6] Su played a "critical role"[7] in developing the "recipe"[2] to make copper connections work with semiconductor chips instead of aluminum, "solving the problem of preventing copper impurities from contaminating the devices during production".[7] Working with various IBM design teams on the details of the device, Su explained, "my specialty was not in copper, but I migrated to where the problems were".[6] The copper technology was launched in 1998,[7] resulting in new industry standards[22] and chips that were up to 20% faster than the conventional versions.[6][7]
AMD was close to ruin when Lisa took over. It had completely lost the graphics and x64 wars and was limping by on low margin games consoles.
Since then Epyc has broken Intel. Anyone buying Xeon's today is at credible risk of being fired over it when someone notices the power bill relative to their competition.
The graphics ramp is behind Nvidia in mind share and market share. The ROCm software stack gets a lot of grief on here. Nevertheless, Nvidia lost the frontier and el cap bids and now Microsoft is running GPT4 on AMD hardware. Sounds pretty good to me given it's one product line of many.
If turning a multinational semiconductor firm away from the edge of bankruptcy and into the profitable conglomerate it is today doesn't qualify as "solving a hard problem" I don't know what would. It's way beyond what I'd be capable of doing.
Major clouds have been investing in ARM alternatives.
x86 is still king compatibility matters a lot. But it is not as simple as Su paints for teams of x86 chip designers to switch to arm chip designers and reach top place, specifically because fabs also matter and AMD is the player in the market with less money to pay the fabs.
GPU market would be hard to recover, and reason to use AMD (for the money printing AI) is just budget. Software is not good enough, it’s understandably not in AMD’s DNA, it was simply lacking in budget as it was close to bankruptcy when CUDA started taking off.
emphasis on top, ofc great designers would still design great no matter the ISA, but best and great is different
It says in the article that he was asked by readers to do a Lisa Su interview. The headline is a little misleading, they don't talk much about technological problems. The interview is a soft light tour of her career and a few attempts to get her to talk about the present day business. Interviews with someone like Musk or Jensen are much more technically intense.
Honestly this interview feels bearish for AMD. Su's performance is not good. Thompson repeatedly pushes her to reflect on past mistakes, but it's just not happening. The reason why AMD has fallen so far behind NVIDIA shine through clear as day and it doesn't look like it's going to get fixed anytime soon.
Su's problem and therefore AMD's problem is that she doesn't want to think about software at all. Hardware is all she knows and she states that openly. Nor does she seem to consider this a weakness. The problem goes back to the very start of her career. The interview opens with Thompson saying she faced a choice between computer science and electronics engineering at MIT, and she picked EE because it was harder. Is that true? She's nowhere in AI due to lack of sufficient skilled devs so now would be a good time to talk up the importance of software, but no, she laughs and says sure! CS seemed easy to her because you "just" write software instead of "building things", whereas in electronics your stuff "has to work". End of answer.
He tries to get a comment on the (in hindsight) not great design tradeoffs made by the Cell processor, which was hard to program for and so held back the PS3 at critical points in its lifecycle. It was a long time ago so there's been plenty of time to reflect on it, yet her only thought is "Perhaps one could say, if you look in hindsight, programmability is so important". That's it! In hindsight, programmability of your CPU is important! Then she immediately returns to hardware again, and saying how proud she was of the leaps in hardware made over the PS generations.
He asks her if she'd stayed at IBM and taken over there, would she have avoided Gerstner's mistake of ignoring the cloud? Her answer is "I don’t know that I would’ve been on that path. I was a semiconductor person, I am a semiconductor person." - again, she seems to just reject on principle the idea that she would think about software, networking or systems architecture because she defines herself as an electronics person.
Later Thompson tries harder to ram the point home, asking her "Where is the software piece of this? You can’t just be a hardware cowboy ... What is the reticence to software at AMD and how have you worked to change that?" and she just point-blank denies AMD has ever had a problem with software. Later she claims everything works out of the box with AMD and seems to imply that ROCm hardly matters because everyone is just programming against PyTorch anyway!
The final blow comes when he asks her about ChatGPT. A pivotal moment that catapults her competitor to absolute dominance, apparently catching AMD unaware. Thompson asks her what her response was. Was she surprised? Maybe she realized this was an all hands to deck moment? What did NVIDIA do right that you missed? Answer: no, we always knew and have always been good at AI. NVIDIA did nothing different to us.
The whole interview is just astonishing. Put under pressure to reflect on her market position, again and again Su retreats to outright denial and management waffle about "product arcs". It seems to be her go-to safe space. It's certainly possible she just decided to play it all as low key as possible and not say anything interesting to protect the share price, but if I was an analyst looking for signs of a quick turnaround in strategy there's no sign of that here.
In my point of view AMD is going down not because nVidia, but because of ARM and Qualcomm. AMD Ryzen x64 cash cow is going to start declining soon both in the server and consumer space.
I saw this clear as day when M1 Macbooks came out and Amazon AWS Graviton servers becoming more popular and cheaper. It was inevitable that the PC world was going to move to ARM soon, in fact I am surprised that it took this long to get viable ARM PC laptops (only this year).
So unless AMD has some secret ARM or RISC-V research division close to launch a product I don't see how it is going to survive long term.
Electrical engineers do generally think software is easy. Even when their day is a horror show of TCL and verilog. In fairness I think hardware is horrendously difficult, so maybe they're not wrong.
I was todays years old, when I found out that Lisa Su (CEO of AMD) and Jensen Huang (co-founder and CEO of NVIDIA) are relatives! If you can't do a merge, it's good to have family onboard ;-)
They didn't know about their relation until much later, but if they had Jensen would have been the "cousin you don't want to be like" - he went to Oregon State and worked at a Denny's while Lisa Su went to Bronx Science and on to MIT.
> he went to Oregon State and worked at a Denny's while Lisa Su went to Bronx Science and on to MIT.
Seems like a bad vibe to imply someone shouldn't aspire to go to state school or work a humble job for money to get through it, even though given both options, indeed they may dream about the fancy one. Denny's has the best milkshakes anyway and state school is probably a much more sensible place to attend.
2 replies →
Although he did graduate high school two years early, so he had the intuition. Maybe his parents thought working food service for a bit was a rite of passage.
3 replies →
The "cousin you don't want to be like" is someone with an electrical engineering degree at 21?
1 reply →
And then he went to Stanford so hardly a failure...
3 replies →
(-; Indeed always warming to the core to see productive emulation between parallel lines! I'm sure after all their achievements, neither of them wastes time pondering woulda, shoulda, cuda…
You made my day with woulda shoulda cuda.
I’m going to watch finding NeMo now
I couldn't believe this so I had to Google it. First cousins once removed. I don't really know what to think about this...
All US presidents are directly related to each other: https://curiousmindmagazine.com/all-us-presidents-including-...
The world's a stage. :)
28 replies →
> please don’t think of AMD as an x86 company, we are a computing company, and we will use the right compute engine for the right workload.
Love that quote.
Yep, I think that's the best quote of the interview and the runner up is the Napster quote.
People seems to have forgotten that Intel used to had StrongARM in their lineup, with the same logic Intel is a computing company not x86 company, similar to AMD [1].
For one of the latest trends in the computing landscape please check this new Embedded+ platform recently introduced by AMD [2]. I'm biased towards IoT but I've a very strong feeling that IoT is the killer application of AI, similar to AI is a killer application of HPC as the latter statement was mentioned by Lisa Su during the interview. The turning point for IoT however is yet to happen, it will probably happened when the number of machine nodes talking to directly to each other, surpassing the number of machine nodes talking to human, then it will be its ChatGPT moment.
[1] StrongARM:
https://en.wikipedia.org/wiki/StrongARM
[2] AMD Unveils Embedded+ Architecture; Combines Embedded Processors with Adaptive SoCs to Accelerate Time-to-Market for Edge AI Applications:
https://www.amd.com/en/newsroom/press-releases/2024-2-6-amd-...
Crazy that they need to write such things because incompetent investors think that ISA has huge implications
It has huge implications because it makes competing in that market much harder due to licensing issues. Intel and AMD have a duopoly on the x86 market which compromises a huge chunk of server and personal computing, but that is changing fast. If they go ARM (or risc-v or whatever) they will have more competition to contend with, including their existing cloud computing clients designing their own chips and fabbing them with other foundries.
2 replies →
The shocker is: a lot of engineers think that the ISA has huge implications.
4 replies →
ISAs do have huge implications. A poorly designed ISA can balloon the instruction count, which has a direct impact on program size and the call stack.
1 reply →
It is good to mention the AMD's Steam Deck CPU [1] running in the Steam Deck [2] and not less important that the Steam Deck also has Linux (and KDE) incorporated [3].
[1] https://www.techpowerup.com/cpu-specs/steam-deck-cpu-lcd.c33...
[2] https://store.steampowered.com/steamdeck
[3] https://help.steampowered.com/en/faqs/view/671A-4453-E8D2-32...
I have endless admiration for Lisa Su, but lets be honest, the reason AMD and Nvidia are so big today is that Intel has had amazingly bad management since about 2003.
They massively botched the 32-bit to 64-bit transition....I did some work for the first Itanium, and everybody knew even before it came out that it would never show a profit. And they were contractually obligated to HP to not make 64-bit versions of the x86....so we just had to sit there and watch while AMD beat us to 1 Gigahertz, and had the 64-bit x86 market to itself....
When they fired Pat Gelsinger, their doom was sealed. Thank God they hired him back, but now they are in the same position AMD and Nvidia used to be in: Intel just has to wait for Nvidia and AMD to have bad management for two straight decades....
You're talking about very old events. Intel made mistakes during 2000-2005 (which allowed the rise of Opteron) but they crushed it from 2006-2016. Then they had different problems from 2016-2024.
Intel went into Rest & Vest mode after Haswell (4000 series) in 2013 - they hardly improved until 2020 - 6 lost generations of relabeled CPUs like 13000 - 14000 series (lots of money paid for innovative new igpu marketing names like 520 and 620!). They found out that community college talent does not make good employees in the C-Suite ... The iris pro 5200 (2013) wasn't improved upon until 2020 !
2 replies →
Thanks for making me feel old :-)
There was a palpable change when Andy Grove retired. It was like the elves leaving Middle earth.
I don't mean to take away from Intel's underwhelming management.
But regardless, Keller's Athlon 64 or Zen are great competitors.
Likewise, CUDA is Nvidia's massive achievement. The growth strategy of that product (involving lots of free engineer hours given to clients on-site) deserves credit.
// I don't mean to take away from Intel's underwhelming management
chuckle lets give full credit where credit is due :-)
Athlon was an epochal chip. Here's the thing though---if you are a market leader, one who was as dominant as Intel was, it doesn't matter what the competition does, you have the power to keep dominating them by doing something even more epochal.
That's why it can be so frustrating working for a #2 or #3 company....you are still expected to deliver epochal results like clockwork. But even if you do, your success is completely out of your hands. Bringing out epochal products doesn't get you ahead, it just lets you stay in the game. Kind of like the Red Queen in Alice in Wonderland. You have to run as fast as you can just to stay still.
All you can do is try to stay in the game long enough until the #1 company makes a mistake. If #1 is dominate enough, they can make all kinds of mistakes and still stay on top, just by sheer market inertia. Intel was so dominate that it took DECADES of back-to-back mistakes to lose its dominate position.
Intel flubbed the 32-64 bit transition. On the low end, it flubbed the desktop to mobile transition. On the high end, it flubbed the CPU-GPU transition.
Intel could have kept its dominate position if it had only flubbed one of them. But from 2002 to 2022, Intel flubbed every single transition in the market.
Its a measure of just how awesome Intel used to be that it took 20 years....but there's only so many of those that you can do back-to-back and still stay #1.
I began supporting AMD as my choice for a gaming CPU when they began trouncing Intel in terms of performance vs. total draw power with an attractive price point around 2016 or so.
Then, a wave of speculative execution vulnerabilities were discovered / disclosed, resulting in an even larger differential for performance and power use after the SPECTRE patches were applied.
Considering this, I'm not sure that it's fair to cast the successes of AMD as mere failures from Intel. Dr. Su is simply a brilliant engineer and business leader.
Well geez, Intel might have amazingly bad management but where on the scale do we put AMD not giving any fucks about challenging the CUDA monopoly for going on 10 years now?
Instead they put mediocre people on things like OpenCL which even university students forced to use the mess could tell was going nowhere.
AMD basically did not have any money from 2010 until 2020. They were teetering near death. No money for r&d no money for new features in their gpus no money no money no money. There excavator architecture was extremely foolish trying to support multiple instruction dispatches with only 1 ALU and FPU per core! This was corrected with Ryzen and the last 4 years they have been able to pay off their massive debts, especially with the 5000 series!
Going forward I expect we will see much more innovation from them because now they have some spare cash to spend on real innovations hopefully in hardware AND software! Note that Intel is much worse than AMD in software!
I have endless admiration for Lisa Su, but lets be honest, the reason Nvidia is so big today is that AMD has had amazingly bad management.
Was I the only one who was expecting Lisa Su to attempt Leetcode hard in this interview?
It is an interview with a CEO - expect platitudes and fortune cookie wisdom that contradict their own (corporate) behaviours.
As an IC this is exactly what makes 'fireside chats' with executives so unsatisfying.
3 replies →
Not really. I wouldn't classify leetcode as "hard problems". Maybe dumb problems that don't really help anyone, but no, not "hard problems"
Sounds like you have beef with leetcode. I think that comment was referencing the problems on leetcode tagged with the "hard" label.
This article inspired me to check out the respective market caps for Intel and AMD. Hell of a turnaround! I remember the raging wars between the Pentiums and the Athlons. Intel won. The GPU wars between Nvidia and ATI; Nvidia won. Thereafter AMD, the supposed loser to Intel, absorbed ATI, the supposed loser to Nvidia. But I love that the story didn't end there. Look at what AMD did with Sony PlayStation (extensively discussed in this interview)...and that's without getting into the contemporary GPU AI revolution that's likely driving AMD's $250+ billion market cap. Epic!
And yet they still can't solve the problem of their GPU driver/software stack for ML being much worse than NVidia's. It seems like the first step is easy: pay more for engineers. AMD pays engineers significantly less than NVidia, and it's presumably quite hard to build a competitive software stack while paying so much less. You get what you pay for.
Everyone does software poorly, hardware companies more so.
Well this is glaringly obvious to whole world, and Nvidia managed to get it right. Surely a feat that can be repeated elsewhere when enough will is spread over some time. And it would make them grow massively, something no shareholder ever frowns upon.
4 replies →
Exactly. ROCm is only available for the top tier RX 7900 GPUs and you’re expected to run Linux.
AMD was (is) tinkering with a translation layer for CUDA, much like how WINE translates directX. Great idea but it’s been taking a while in this fast paced market.
> AMD was (is) tinkering with a translation layer for CUDA
From what I understand, they dropped the contract with the engineer who was working on it.
Fortunately, as part of the contract, said engineer stipulated that the project would become open source, so now it is, and is still being maintained by that engineer.
> ROCm is only available for the top tier RX 7900 GPUs and you’re expected to run Linux.
Fixed it for you: ROCm is only officially supported for the top tier RX 7900 GPUs and you’re expected to run Linux.
Desktop class cards work if you apply a "HSA version override".
1 reply →
Came here to say this. They only just recently got an AMD GPU on MLPerf thanks to a (different company), Tinycorp by George Hotz. I guess basic ML performance is too hard a problem.
I dunno, it a world where hardware companies like, sold hardware, and then software companies wrote software and sold that could be pretty nice. It is cool that Hotz is doing something other than contribute to an anticompetitive company’s moat.
A couple of thoughts here.
* AMD's traditional target market for its GPUs has been HPC as opposed to deep learning/"AI" customers.
For example, look at the supercomputers at the national labs. AMD has won quite a few high profile bids with the national labs in recent years:
- Frontier (deployment begun in 2021) (https://en.wikipedia.org/wiki/Frontier_(supercomputer)) - used at Oak Ridge for modeling nuclear reactors, materials science, biology, etc.
- El Capitan (2023) (https://en.wikipedia.org/wiki/El_Capitan_(supercomputer)) - Livermore national lab
AMD GPUs are pretty well represented on the TOP500 list (https://top500.org/lists/top500/list/2024/06/), which tends to feature computers used by major national-level labs for scientific research. AMD CPUs are even moreso represented.
* HPC tends to focus exclusively on FP64 computation, since rounding errors in that kind of use-case are a much bigger deal than in DL (see for example https://hal.science/hal-02486753/document). NVIDIA innovations like TensorFloat, mixed precision, custom silicon (e.g., the "transformer engine") are of limited interest to HPC customers. It's no surprise that AMD didn't pursue similar R&D, given who they were selling GPUs to.
* People tend to forget that less than a decade ago, AMD as a company had a few quarters of cash left before the company would've been bankrupt. When Lisa Su took over as CEO in 2014, AMD market share for all CPUs was 23.4% (even lower in the more lucrative datacenter market). This would bottom out at 17.8% in 2016 (https://www.trefis.com/data/companies/AMD,.INTC/no-login-req...).
AMD's "Zen moment" didn't arrive until March 2017. And it wasn't until Zen 2 (July 2019), that major datacenter customers began to adopt AMD CPUs again.
* In interviews with key AMD figures like Mark Papermaster and Forrest Norrod, they've mentioned how in the years leading up to the Zen release, all other R&D was slashed to the bone. You can see (https://www.statista.com/statistics/267873/amds-expenditure-...) that AMD R&D spending didn't surpass its previous peak (on a nominal dollar, not even inflation-adjusted, basis) until 2020.
There was barely enough money to fund the CPUs that would stop the company from going bankrupt, much less fund GPU hardware and software development.
* By the time AMD could afford to spend on GPU development, CUDA was the entrenched leader. CUDA was first released in 2003(!), ROCm not until 2016. AMD is playing from behind, and had to make various concessions. The ROCm API is designed around CUDA API verbs/nouns. AMD funded ZLUDA, intended to be a "translation layer" so that CUDA programs can run as a drop-in on ROCm.
* There's a chicken-and-egg problem here.
1) There's only one major cloud (Azure) that has ready access to AMD's datacenter-grade GPUs (the Instinct series).
2) I suspect a substantial portion of their datacenter revenue still comes from traditional HPC customers, who have no need for the ROCm stack.
3) The lack of a ROCm developer ecosystem means that development and bug fixes come much slower than they would for CUDA. For example, the mainline TensorFlow release was broken on ROCm for a while (you had to install the nightly release).
4) But, things are improving (slowly). ROCm 6 works substantially better than ROCm 5 did for me. PyTorch and TensorFlow benchmark suites will run.
Trust me, I share the frustration around the semi-broken state that ROCm is in for deep learning applications. As an owner of various NVIDIA GPUs (from consumer laptop/desktop cards to datacenter accelerators), in 90% of cases things just work on CUDA.
On ROCm, as of today it definitely doesn't "just work". I put together a guide for Framework laptop owners to get ROCm working on the AMD GPU that ships as an optional add-in (https://community.frame.work/t/installing-rocm-hiplib-on-ubu...). This took a lot of head banging, and the parsing of obscure blogs and Github issues.
TL;DR, if you consider where AMD GPUs were just a few years ago, things are much better now. But, it still takes too much effort for the average developer to get started on ROCm today.
Summary: AMD works if you spend 500m USD+ with them. Then they'll throw an army of their own software engineers into the contract who will hold your hand every step of the way, and remove all the jank for you. By contrast, since at least 10 years ago, I could buy any GTX card and CUDA worked out of the box, and that applied right down to a $99 Jetson Nano.
AMD's strategy looks a lot like IBM's mainframe strategy of the 80s. And that didn't go well.
9 replies →
Small correction: CUDA was first released in 2007 and of course Nvidia was also aiming at HPC before the AlexNet moment.
Good summary. There was also the 2010's multivendor HSA and OpenCL software evolution directions that ended up losing other vendors on the way and many customers turned out to accept the proprietary Cuda.
And yet people seem to work just fine with ML on AMD GPUs when they aren’t thinking about Jensen.
I have a 7900 XTX. There's a known firmware crash issue with ComfyUI. It's been reported like a year ago. Every rocm patch release I check the notes, and every release it goes unfixed. That's not to go into the intense jank that is the rocm debian repo. If we need DL at work, I'll recommend Nvidia, no question.
Which AMD GPUs? Most consumer AMD GPUs don't even support ROCm.
11 replies →
Omg. I know this is mostly marketing speaking, but this is her reply when asked about AMD's reticence to software:
> Well, let me be clear, there’s no reticence at all. [...] I think we’ve always believed in the importance of the hardware-software linkage and really, the key thing about software is, we’re supposed to make it easy for customers to use all of the incredible capability that we’re putting in these chips, there is complete clarity on that.
I'm baffled how clueless these CEOs sometimes seem about their own product. Like, do you even realize that this the reason why Nvidia is mopping the floor with your stuff? Have you ever talked to a developer who had to work with your drivers and stack? If you don't start massively investing on that side, Nvidia will keep dominating despite their outrageous pricing. I really want AMD to succeed here, but with management like that I'm not surprised that they can't keep up. Props to the interviewer for not letting her off the hook on this one after she almost dodged it.
What is she supposed to say? Perhaps "our products have bad software, don't buy them, go buy Nvidia instead"?
She could admit that they fell behind on this one and really need to focus on closing the gap now. But instead she says it's all business as usual, which assures me that I won't give their hardware another shot for quite a while.
6 replies →
She could say "We are in a compute gold rush and yet AMD's stock didn't gain anything in the last 6 months, so I hereby submit my resignation". That would work.
1 reply →
I'd love to know if any domain experts have a write up on what the the talent+ time+financial investment it would take for AMD to come up with with something that is a worthy rival to CUDA. Very curious to understand what the obstacles are.
~5 years. Medium-sized team in-house + hordes (hundreds, thousands) of engineers in the field helping clients on-site, writing code for them directly upstreamed to drivers, core libs, etc. (iteratively optimized in-house, ship feature, rinse and repeat). Story of the PlayStation SDKs, of DX too, but above all CUDA (they really outdid this strategy), now for cuDNN and so much more.
It takes incompressible time because you have to explore the whole space, cover most bases; and it takes an industry several years (about one "gen" / hardware cycle) to do that meaningfully.It helps when your platform is disruptive and customers move fast.
Maybe 3 years at best if you start on a new ideal platform designed for it from scratch. And can throw ungodly amount of money fast at it (think 5K low-level engineers roaming your installed base).
Maybe 10+ yrs (or never) if you're alone, poor, and Radeon (j/k but to mean it's non-trivial).
I’d say it mainly needs persistence and good execution (library support). NVIDIA has co-developed CUDA with their hardware, and largely stayed compatible with it, since around 2009, and around 2012 it first started taking off in the HPC space. Years later this enabled first their boom in crypto and then an even bigger one in AI. I don’t think this amount of R&D would be out of reach of today’s AMD (as NVIDIA wasn’t any bigger back then), but the backing of it needs to come from the very top.
First, they need to work with kernel devs to finally fix their drivers. Like, Nvidia used to be a "pain in the ass" here as well (that's a literal quote from Torvalds), so simply by contributing more than nothing, they could have taken the lead. But they definitely screwed this one up.
Second, they need to fix their userspace stack. ROCm being open source and all is great in principle, but simply dropping your source to the masses doesn't make it magically work. They need to stop letting it linger by either working with the open source community (huge time investment) or do it themselves (huge money investment).
4 replies →
I wonder if they really need a CUDA rival.
This AI stuff has progressed a bit. Intel has been working on interesting stuff with OneAPI. It might be the case that things have progressed to the point where the primitives are well enough understood that you need something more like a good library rather than a good compiler.
In the end, more people seem to love BLAS than Fortran, after all.
2 replies →
I don't want a CUDA rival. I want to get the entire pile of CUDA code that is already written and run it on AMD GPUs without any kind of tweak or rewrite, and have it just work every time
Compatibility with existing code is very important. People can't afford to rewrite their stuff just to support AMD, and thus they don't
AMD is kind of trying to do this with rocm and HIP, but whatever they are doing it's not enough
My theory is that someone came up with the bright idea of allowing more open source in the stack and that that would allow them to get it all done via crowd sourcing and on the cheap. But if true it was a quite naive view of how it might work.
If instead they said let's take the money we should invested in internal development and build an open developer community that will leverage our hardware to build a world class software stack it might have been a little better.
1 reply →
I spotted this recent post https://www.reddit.com/r/LocalLLaMA/comments/1deqahr/comment... that was pretty interesting:
> When I was working on TVM at Qualcomm to port it to Hexagon a few years ago we had 12 developers working on it and it was still a multiyear long and difficult process.
> This is also ignoring the other 20 or so developers we had working on Hexagon for LLVM, which did all of the actual hardware enablement; we just had to generate good LLVM IR. You have conveniently left out all of the LLVM support that this all requires as AMD also uses LLVM to support their GPU architectures.
> Funny enough, about a half dozen of my ex coworkers left Qualcomm to go do ML compilers at AMD and they're all really good at it; way better than I am, and they haven't magically fixed every issue
> It's more like "hire 100 additional developers to work on the ROCM stack for a few years"
This last statement sounds about right. Note that ROCm has over 250 repos on Github, a lot of them pretty active: https://github.com/orgs/ROCm/repositories?type=all - I'm sure an enterprising analyst who was really interested could look at the projects active over the past year and find unique committers. I'd guess it's in the hundreds already.
I think if you click through the ROCm docs https://rocm.docs.amd.com/en/latest/ (and maybe compare to the CUDA docs https://docs.nvidia.com/cuda/ ) you might get a good idea of the differences. ROCm has made huge strides over the past year, but to me, the biggest fundamental problem is still that CUDA basically runs OOTB on every GPU that Nvidia makes (with impressive backwards and in some cases even forwards compatibility to boot https://docs.nvidia.com/deploy/cuda-compatibility/ ) on both Linux and Windows, and... ROCm simply doesn't.
I think the AMD's NPUs complicate things a bit as well. It looks like it's its currently running on its own ONNX/Vitis (Xilinx) stack https://github.com/amd/RyzenAI-SW , and really it should either get folded into ROCm (or a new SYCL/oneAPI-ish layer needs to be adopted to cover everything).
> is a worthy rival to CUDA
Vulkan Compute already exists
But when Developers still continue buying NVIDIA for CUDA, because developers only target CUDA for their applications it is a chicken-egg scenario, similar to Linux vs Windows.
> Like, do you even realize that this the reason why Nvidia is mopping the floor with your stuff?
Are they?
Generally, my experience has been that AMD products generally just work even if they're a bit buggy sometimes, while Nvidia struggles to get a video signal at all.
Seems perfectly reasonable to focus on what matters, while Nvidia is distracted by the AI fad.
I'm not talking about gaming, I'm talking about general purpose computing (although even for gaming your statement is pretty bold). Since she's CEO of a publicly traded company, it seems pretty weird that she would ignore the fields where the money is at, while Nvidia becomes the most valuable company in the world. So she's not just ignoring developers' wants but also her stockholders'.
6 replies →
even if the LLM thing is an "AI fad" there are many other things that ML is used for that matter (to the people spending real money on GPUs - think A6000, H100, not gamer cards)
2 replies →
I’m not anti-CEO, I think they play an important role, but why would you interview a CEO about hard technological problems?
She worked as a researcher in the field for decades. Moore's law only kept up because of some of the techniques she developed.
> During her time at IBM,[6] Su played a "critical role"[7] in developing the "recipe"[2] to make copper connections work with semiconductor chips instead of aluminum, "solving the problem of preventing copper impurities from contaminating the devices during production".[7] Working with various IBM design teams on the details of the device, Su explained, "my specialty was not in copper, but I migrated to where the problems were".[6] The copper technology was launched in 1998,[7] resulting in new industry standards[22] and chips that were up to 20% faster than the conventional versions.[6][7]
AMD was close to ruin when Lisa took over. It had completely lost the graphics and x64 wars and was limping by on low margin games consoles.
Since then Epyc has broken Intel. Anyone buying Xeon's today is at credible risk of being fired over it when someone notices the power bill relative to their competition.
The graphics ramp is behind Nvidia in mind share and market share. The ROCm software stack gets a lot of grief on here. Nevertheless, Nvidia lost the frontier and el cap bids and now Microsoft is running GPT4 on AMD hardware. Sounds pretty good to me given it's one product line of many.
If turning a multinational semiconductor firm away from the edge of bankruptcy and into the profitable conglomerate it is today doesn't qualify as "solving a hard problem" I don't know what would. It's way beyond what I'd be capable of doing.
Major clouds have been investing in ARM alternatives. x86 is still king compatibility matters a lot. But it is not as simple as Su paints for teams of x86 chip designers to switch to arm chip designers and reach top place, specifically because fabs also matter and AMD is the player in the market with less money to pay the fabs.
GPU market would be hard to recover, and reason to use AMD (for the money printing AI) is just budget. Software is not good enough, it’s understandably not in AMD’s DNA, it was simply lacking in budget as it was close to bankruptcy when CUDA started taking off.
emphasis on top, ofc great designers would still design great no matter the ISA, but best and great is different
Lisa Su is one of the (surprisingly rare) tech CEOs who comes from an engineering background.
You mean like the CEOs of Intel, Nvidia, TSMC, ASML, Micron etc?
10 replies →
And so is her first cousin (once removed) Jensen Huang of Nvidia!
Much of the hard problems these days require scale and with it coordination
It says in the article that he was asked by readers to do a Lisa Su interview. The headline is a little misleading, they don't talk much about technological problems. The interview is a soft light tour of her career and a few attempts to get her to talk about the present day business. Interviews with someone like Musk or Jensen are much more technically intense.
Honestly this interview feels bearish for AMD. Su's performance is not good. Thompson repeatedly pushes her to reflect on past mistakes, but it's just not happening. The reason why AMD has fallen so far behind NVIDIA shine through clear as day and it doesn't look like it's going to get fixed anytime soon.
Su's problem and therefore AMD's problem is that she doesn't want to think about software at all. Hardware is all she knows and she states that openly. Nor does she seem to consider this a weakness. The problem goes back to the very start of her career. The interview opens with Thompson saying she faced a choice between computer science and electronics engineering at MIT, and she picked EE because it was harder. Is that true? She's nowhere in AI due to lack of sufficient skilled devs so now would be a good time to talk up the importance of software, but no, she laughs and says sure! CS seemed easy to her because you "just" write software instead of "building things", whereas in electronics your stuff "has to work". End of answer.
He tries to get a comment on the (in hindsight) not great design tradeoffs made by the Cell processor, which was hard to program for and so held back the PS3 at critical points in its lifecycle. It was a long time ago so there's been plenty of time to reflect on it, yet her only thought is "Perhaps one could say, if you look in hindsight, programmability is so important". That's it! In hindsight, programmability of your CPU is important! Then she immediately returns to hardware again, and saying how proud she was of the leaps in hardware made over the PS generations.
He asks her if she'd stayed at IBM and taken over there, would she have avoided Gerstner's mistake of ignoring the cloud? Her answer is "I don’t know that I would’ve been on that path. I was a semiconductor person, I am a semiconductor person." - again, she seems to just reject on principle the idea that she would think about software, networking or systems architecture because she defines herself as an electronics person.
Later Thompson tries harder to ram the point home, asking her "Where is the software piece of this? You can’t just be a hardware cowboy ... What is the reticence to software at AMD and how have you worked to change that?" and she just point-blank denies AMD has ever had a problem with software. Later she claims everything works out of the box with AMD and seems to imply that ROCm hardly matters because everyone is just programming against PyTorch anyway!
The final blow comes when he asks her about ChatGPT. A pivotal moment that catapults her competitor to absolute dominance, apparently catching AMD unaware. Thompson asks her what her response was. Was she surprised? Maybe she realized this was an all hands to deck moment? What did NVIDIA do right that you missed? Answer: no, we always knew and have always been good at AI. NVIDIA did nothing different to us.
The whole interview is just astonishing. Put under pressure to reflect on her market position, again and again Su retreats to outright denial and management waffle about "product arcs". It seems to be her go-to safe space. It's certainly possible she just decided to play it all as low key as possible and not say anything interesting to protect the share price, but if I was an analyst looking for signs of a quick turnaround in strategy there's no sign of that here.
In my point of view AMD is going down not because nVidia, but because of ARM and Qualcomm. AMD Ryzen x64 cash cow is going to start declining soon both in the server and consumer space.
I saw this clear as day when M1 Macbooks came out and Amazon AWS Graviton servers becoming more popular and cheaper. It was inevitable that the PC world was going to move to ARM soon, in fact I am surprised that it took this long to get viable ARM PC laptops (only this year).
So unless AMD has some secret ARM or RISC-V research division close to launch a product I don't see how it is going to survive long term.
12 replies →
Electrical engineers do generally think software is easy. Even when their day is a horror show of TCL and verilog. In fairness I think hardware is horrendously difficult, so maybe they're not wrong.
7 replies →