layperson with no industry knowledge, but it seems like nvidia's CUDA moat will fall in the next 2-5 years. It seems impossible to sustain those margins without competition coming in and getting a decent slice of the pie
replies(5):
You are right, eventually something's gotta give. The path for this next leg isn't yet apparent to me.
P.s. how much is an exaflop or petaflop, and how significant is it? The numbers thrown around in this article don't mean anything to me. Is this new cluster way more powerful than the last top?
Most customers care about cost-effectiveness more than best-in-class raw-performance, a fact that AMD has ruthlessly exploited over the past 8 years. It helps that AMD products are occasionally both.
> El Capitan – we don’t yet know how big of a portion yet as we write this – with 43,808 of AMD’s “Antares-A” Instinct MI300A devices
By comparison XAI announced that they have 100k H100s. MI300A and H100s have roughly similar performance. Meta says they're training on more than 100k H100s for Llama-4, and have the equivalent of 600k H100s worth of compute. (Note that compute and networking can be orthogonal).
Also, Nvidia B200s are rolling out now. They offer 2-3x the performance of H100s.
I know a lot developing on apples silicon and just pushing it to clusters for bigger runs. So why not run it on an apple GPU there?
The MI300A can get to 150% the FP64 peak performance that B200 devices can get, although AMD GPUs have historically underperformed their spec more than Nvidia GPUs. It's possible that B200 devices are actually behind for HPC.
Also, of the top 10, AMD has 5 systems.
1 petaflop = 10^15 flops = 1,000,000,000,000,000 flops.
1 exaflop = 10^18 flops = 1,000,000,000,000,000,000 flops.
Note that these are simply powers of 10, not powers of 2, which are used for storage for example.
It really is that hard, and the fabrication side of the issue the easy part from Nvidia's perspective - you just pay TSMC a shitload of money. Nvidia's real victory (besides leading on performance-per-watt) is that their software stack doesn't suck. They invested in complex shader units and tensor accelerators that scale with the size of the card rather than being restrained in puny and limited NPUs. CUDA unified this featureset and was industry-entrenched for almost a decade, which gave it pretty much any feature you could want be it crypto acceleration or AI/ML primitives.
The ultimate tragedy is that there was a potential future where a Free and Open Source CUDA alternative existed. Apple wrote the OpenCL spec for exactly that purpose and gave it to Khronos, but later abandoned it to focus on... checks clipboard MLX and Metal Performance Shaders. Oh, what could have been if the industry weren't so stingy and shortsighted.
It does seem like Nvidia is prioritizing int8 / fp8 performance over FP64, which given the current state of the ML marketplace is a great idea.
For everything that isn't machine learning, I frankly feel like it's the other way around. Apple's "solution" to these edge cases is telling people to write compute shaders that you could write in Vulkan or DirectX instead. What sets CUDA apart is an integration with a complex acceleration pipeline that Apple gave up trying to replicate years ago.
When cryptocurrency mining was king-for-a-day, everyone rushed out to buy Nvidia hardware because it supported accelerated crypto well from the start. The same thing happened with the AI and machine learning boom. Apple and AMD were both late to the party and wrongly assumed that NPU hardware would provide a comparable solution. Without a CUDA competitor, Apple would struggle more than AMD to find market fit.
Nominally, a measurement in "flops" is how many (typically 32-bit) FLoating-point Operations Per Second the hardware is capable of performing, so it's an approximate measure of total available computing power.
A high-end consumer-grade CPU can achieve on the order of a few hundred gigaflops (let's say 250, just for a nice round number). https://boinc.bakerlab.org/rosetta/cpu_list.php
A petaflop is therefore about four thousand of those; multiply by another thousand to get an exaflop.
For another point of comparison, a high-end GPU might be on the order of 40-80 teraflops. https://www.tomshardware.com/reviews/gpu-hierarchy,4388-2.ht...
NVidia currently has 80-90% gross margins on their LLM GPUs, that’s all the incentive another company needs to invest money into a CUDA alternative.
Which doesn’t help with understanding how much more impressive these are than the last clusters, but does to me at least put the amount of compute these clusters have into focus.
If you look at the table toward the bottom, no matter how you slice it, Nvidia has 50% of the total cores, 50% of the total flops, and 90% of the total systems among the Top 500, while AMD has 26% of the total cores, 27.5% of the total flops, and 7% of the total systems.
Is it a matter of newly-added compute?
> This time around, on the November 2024 Top500 rankings, AMD is the big winner in terms of adding capacity to the HPC base.
Nvidia's lead is not only cemented by dense silicon. Their designs are extremely competitive, perhaps even a generational leap over what their competitors offer.
For ML, not for HPC. ML and HPC are two completely different, only loosely related fields.
ML tasks are doing great with low precision, 16 and 8 bit precision is fine, arguably good results can be achieved even with 4 bit precision [0][1]. That won't do for HPC tasks, like predicting global weather, computational biology, etc. -- one would need 64 to 128 bit precision for that.
Nvidia needs to decide how to divide the billions of transistors on their new silicon. Greatly oversimplifying, they can choose to make one of the following:
* Card A with *n* FP64 cores, or
* Card B with *2n* FP32 cores, or
* Card C with *4n* FP16 cores, or
* Card D with *8n* FP8 cores, or (theoretically)
* Card E with *16n* FP4 cores (not sure if FP4 is a thing).
Their new offerings are arguably worse than their older offerings for HPC tasks, and the feeling with the HPC crowd is that "Nvidia and AMD are in the process of abandoning this market".
[0] https://papers.nips.cc/paper/2020/file/13b919438259814cd5be8...
My point of reference is that back in undergrad (~10-15 years ago), I recall a class assignment where we had to optimize matrix multiplication on a CPU; typical good parallel implementations achieved about 100-130 gigaflops (on a... Nehalem or Westmere Xeon, I think?).
The code to run these things on apples GPUs exist and is used every day! I don't know anyone using AMD GPUs, but pretty often its nvidia on the cluster and Apple on the laptop. So if nvidia is making these juicy profits, i think apple could seriously think about moving to the cluster if it wants to.
The reason why AMD is behind is that it is behind in hardware. MI300x is more pricey per hour in all the cloud I can find compared to H100, and the MFU is order of magnitude lower compared to NVIDIA for transformers, even though transformers are fully supported. And I get same 40-50% MFU in TPU for the same code. If anyone is investing >10 million dollar for hardware, they sure can invest a million dollar to rewrite everything in whatever language AMD asks them to if it is cheaper.
Which does make the clusters a fair bit less impressive, but also a lot more sensibly sized.
https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvid...
But at these levels of compute, the memory/interconnect bandwidth becomes the bottleneck.
that does a serious disservice to hyperscaler clusters.
You need to develop your own in house solution to distributing workloads.
The difference to regular clusters is that all the memory is globally visible, so machine 0023 can access and modify address 0x0123456789abcdef0123456789abcdef which happens to be on machine 0999.
YES!! Thank you!
> Nvidia's real victory (besides leading on performance-per-watt) is that their software stack doesn't suck
YES! And it's not just CUDA and CUDA-adjacent tools, but also their cuDNN/cuBLAS/etc. libraries. They invest a massive amount of staffing into squeezingt the last drop of performance out of their hardware, identifying areas for improvement and feeding that back to the architects.
> Apple wrote the OpenCL spec for exactly that purpose and gave it to Khronos
Nitpick: Affie Munshi from Apple wrote down a draft and convinced his management to offer it to Khronos, where it was significantly modified over... was it a year or so?... by a number of representatives from a dozen companies or so. A ton of smart people contributed a ton of work into what became the 1.0 version.
And let me tell you that the discussions were often tense, both during the official meetings as well as what happened behind the scenes. The end result was as good as you can expect from a large committee composed of representatives from competing companies.
But, in summary, you get it, unlike so many commenters in HN.
e.g. you could run a H100 at 100% utilization 24/7 for 1 years at $0.4 per kWh (so assuming significant overhead for infrastructure etc.) and that would only cost ~10% of the purchase price of the GPU itself.
From a distance, it kinda sounds like listening to kids brag about their allowance while the adults don't want to talk about their salary, and try to draw wider conclusions from there.
Your other points may be valid.
Also, to be clear, I have no internal info about this, I'm going based on external stuff I've seen.
The practical answer is that all of FAANG will have to pick up the pieces once their supply chain is shattered. Samsung would quickly reach capacity with AMD and potentially Nvidia as priority customers, and Intel will be trying to court Nvidia and Apple as high-margin customers for some low-yield 18A contract. Depending on whether TSMC's Arizona foundry ever reaches operational capacity, they will be balancing orders from Nvidia and Apple in the same way they do today. Given the pitifully low investment, it's not really likely the Arizona facility will make a dent in the supply chain.
Fact is, Nvidia is well positioned to pick up the pieces even if 5nm> processes go away for the next decade. The only question is whether or not people will continue to have demand for CUDA, and the answer has been "yes" since long before crypto and AI were popular. If TSMC was bombed tomorrow, Nvidia would still have demand for their product and they would still have the capacity to sell it. Their competition with AMD would be somewhat normalized and Apple would be blown into the stratosphere upon realizing that they have to contract either Samsung or Intel to stay afloat. The implications for the American economy are a little upsetting but there's nothing particularly world-ending about that scenario. It would be a sad day to be a Geekbench enthusiast but life would go on.
The Pytorch MPS patches are a fun appeasement for developers, but they didn't unthrone Nvidia's demand. They didn't beat Nvidia on performance per watt, they didn't match their price, their scale or CUDA's featureset, and they don't even provide basic server drivers. It's got nothing to do with what brand you prefer and everything to do with what makes actual sense in a datacenter. Apple can't take on Nvidia clusters without copying Nvidia's current architecture - Apple Silicon's current architecture is too inefficient to be a serious replacement to Nvidia clusters.
If Apple wanted to have a shot at entering the cluster game, that window of opportunity closed when Apple Silicon converged on simplified GPU designs. The 2w NPUs and compute shaders aren't going to make Nvidia scared, let alone compete with AMD's market share.
It's no Nvidia but Meta has ordered AMD GPUs.
like do you have actual experience with gov/edu HPC? i doubt it because you couldn't be more wrong - lab HPC clusters are just very very poorly (relative to FAANG) strewn together nodes. there is absolutely no sense in which they are "one single machine" (nothing is "abstracted over" except NFS).
what you're saying is trivially false because no one ever requests all the machines at once (except when they're running linpack to produce top500 numbers). the rest of the time the workflow is exactly like in any industrial cluster: request some machines (through slurm), get those machines, run your job (hopefully you distributed the job across the nodes correctly), release those machines. if i still had my account i could tell you literally how many different jobs are running right now on polaris.
Knowing all of that it still leaves unexplained whether AMD has the needed ~70% of non-accelerated compute (assuming FLOPS) to clear the bar for the headline. It seems unlikely to me... but the article doesn't actually have enough data to be sure one way or the other.
Comparing their crazy expensive custom built HPC to massive arrays of customer grade hardware doesn't bring them additional funds, nor help them more PR wise than being the owner of the fastest individual clusters.
Being at the top of some heap is visibly one of their goal:
Yes, LINPACK is indeed "old" with a heavy focus on compute power. However, its simplicity serves as a reliable baseline for the types of workflows that supercomputers are designed to handle. Also, at their core, most AI workloads perform essentially the same operations as HPC, albeit with less stability—which, I admit, is a feature, but likely the reason AI-focused systems do not prioritize LINPACK as much.
I am simply saying that any useful metric needs to not only be "stable", but also simple to grasp. Take Green500, probably a significant benchmark for understanding how algorithms consume power, but "too complex" to explain: yet, many cloud providers with their AI supercomputers avoid competing against HPC supercomputers in this domain.
This avoidance isn’t necessarily due to secrecy but rather inefficiencies inherent to cloud systems. Consider PUE (Power Usage Effectiveness)—a highly misleading metric that cloud providers frequently tout. PUE can easily be manipulated, especially with the use of liquid cooling, which is why optimizing for it has become a major factor contributing to water disruptions in several large cities worldwide.
Historically, those workloads and users were leading indicators of certain types of things. I don't think that's true anymore. In fact, I wonder if this is mostly a story of the government agencies not being able to compete with the private sector for NVIDIA gpus.
So a 64-bit multiplier is something like 32x more area than a 16-bit multiplier.
But what you say is correct for RAM area or the number of bits you need for register space. So taken holistically, it's difficult to say...
Okay, 64-bit FP is only like 53-bits and 16-bit FP is actually like 11 bits. But you know what I mean. I'm still doing quick napkin math here, nothing formal.
-------
We can ignore adders and subtractor circuits because they are so small. Division is often implemented as reciprocal followed by multiplication circuits for floating point (true division is very expensive).
High precision HPC exists in the private cloud, but you only hear "we don't want to embarrass others" excuses because otherwise you would be able to calculate the cost.
On prem HPC is still very, very much cheaper than hiring out.
And of course there is some serious amount of money sloshing around in this space. Things being hard doesn't mean it's impossible. And there's no shortage of extremely well funded companies working on this stuff. All your favorite trillion $ companies basically. And most of them have their own AI chips too. And probably some reservations about perpetually handing a lot of their cash to Nvidia.
If you want an example of a company that used to have a gigantic moat that is now dealing with a lot of competition, look at Intel. X86 used to be that moat. And that's looking pretty weak lately. One reason that AMD is in the news a lot lately is that they are growing at Intel's expense. Nvidia might be their next target.
i mean rick stevens et al can grab all of polaris too but even so - it's just a bunch of nodes and you're responsible for distributing your work across those nodes efficiently. there's no sense in which it's a "single computer" in any way, shape or form.
We can increase that another 2x and the cost would still be relatively low compared to the price/deprecation of the GPU itself.
Companies like CoreWeave have deployed so many giant clusters (and growing), it is insane. Their IDLE compute is larger than most of the supercompuers out there.
Of course, they aren't on the list either.
My predicition is there will be some strong competition for Nvidia in the coming years.
Since most people use CUDA through some other library (like Torch or TF), I think the dependence on CUDA isn't as strong as you make it seem.
Intel actually has proven to be more clever than AMD in that regard, as DataParalell C++ builds on top of SYCL (it isn't only SYCL), and Intel Fortran now also does GPU offloading.
I feel sorry for you if you believe this. It's not true physically nor is it true on the level of the cache coherence protocol nor is it true from the perspective of the operating system.
Without blindingly fast, otherwise blinding numerical performance dims quite a lot. This is why the Cerebras numbers on heavy numerical problems are competitive up to a pretty severe ceiling. Below that point, their on wafer interconnects suffice, above it they cannot scale the data communications bandwidth necessary.
You are mostly listing irrelevant nice to have things that aren't deal breakers. AMD's consumer GPUs have a long history of being abandoned a year or two after release.
Coupled with Khronos, Intel, AMD never delivering anything comparable with OpenCL, Apple losing interest after Khronos didn't took OpenCL into the direction they wanted, Google never adopting it favouring their Renderscript dialect.
None of them logically relate to another.
One is a question.
And the rest are wrong.
You can't go faster than the speed of light (yet) and traveling a few micrometers will always be much faster than traversing a room (plus routing and switching).
Many HPC tasks nowadays are memory-bound rather than CPU-bound, memory-latency-and-throughput-bound to be more precise. An actual supercomputer would be something like the Cerebras chip, a lot of the performance increase you get is due to having everything on-chip at a given time.
As for Fortran, that doesn't come up much in modern AI stuff. I haven't observed PTX / GCN assembly within AI codebases but maybe you have extra insight there.
That said, I assumed the context of the article was specifically on the topic of AMD GPUs, and not, say, Epyc processors, so if so, it's ultimately irrelevant.
(Also, there are just over 3 exaflops across non-accelerated supers; AMD would need > 2/3 of the remaining share in order to surpass Nvidia on that front as well.)