I've been thinking that NVDA stock is massively overpriced - yes, AI is a hot topic, but their only advantage is the software stack. It is just a matter of time until Intel and AMD realize that they should join hands and do an open-source CUDA alternative for their respecitve GPUs (yes, Intel has competetive GPUs and just like AMD and Nvidia they will try to get a share of the AI chip market share).
Problem is the CUDA advantage is gigantic and it has been known for years in GPGPU processing, way before AI was a meme. AMD has lost countless developers over the year just on hello world style projects. Developers had a solid 6-7 years of living with OpenCL when the green rival had a very mature and nice CUDA sitting there. I’ve been out of that world for a while now, but it was truly painful and turned a lot of devs off programming AMD devices. Now there’s a big moat of entrenched developers that could take decades to displace. It’s like trying to displace C++ with Java 22 — possible, but it’s a slow, slow trudge and everyone still remembers Java 1.4
No, the amount of code written in CUDA for pytorch could easily be rewritten in CUDA for few million or tens of millions of investment. The problem is that it is damn near impossible to get good performance in AMD. For complicated CUDA programs like flash attention(few 100 lines of code), no amount of developers could write those few 100 lines for AMD to get the same performance.
Even worse: GPGPU is not only about LLM or even ML. It's also for computer vision, signal processing, pointcloud processing, e.g. Opencv has backend for CUDA, open3d, PCL the same. Even apple is kind of worse than AMD regarding ecosystem of libraries and open source high performance algorithms - when I tried to port some ICP pipeline to apple metal there was nothing there, most libraries and research code target only CUDA
While I agree with the sentiment towards CUDA, the example is a bit off, given that C++ basically lost all mindshare in distributed computing, to Java and others, and is hardly visible in CNCF projects landscape.
Displacing C++ in compiler development and HFT/HPC/GPGPU with Java 22, most likely not happening, everwhere else it has been loosing mindshare, the current cybersecurity laws versus WG21 attitude towards them, doesn't help.
"AMD is among several companies contributing to the development of an OpenAI-led rival to Cuda, called Triton, which would let AI developers switch more easily between chip providers. Meta, Microsoft and Intel have also worked on Triton."
This is a bit misleading since Triton is a bit higher level than CUDA. But the idea is kind of right - there’s active development of AMD and Intel backends, and Pytorch is investing into Triton as well.
NVDA's moat is over-stated. There are several deep-pocketed players with pretty good AI chips. The big players are training models at such a large scale that they can afford to back them by different architectures. Smaller players use frameworks like Pytorch and Tensorflow, but those are backed by big players buying from Nvidia.
But valuation isn't the NVDA trade right now; it's that there's still a bigger fool.
Different accounting methods from what I gather. The acquisition is being accounted for over a 5 year period for the trailing p/e but not being included in the forward p/e over this 5 year period. This really shows how p/e is not a great metric in a vacuum.
I wish people wouldn't post such pointless comments - the only users who get any value from reading your sentence are people who already share your view and can go "hah yeah!", while you couldn't be bothered to explain why it's your view to anyone who doesn't already think the same thing. Literally no benefit over not saying anything. Sorry to be blunt.
No they don't. Both Intel and AMD compare their newest GPU favorably against Nvidia's H100 that has been on the market longer and soon to be replaced and then it's never H100 NVL for a reason.
Intel and AMD can sell their GPU's only with lower profit margin. If they could match FLOPS per total ownership they would sell much better.
Benchmarks were just run, MI300x is onpar/better than an H100. Next generation of MI (MI325x) is coming out end of the year and those specs look fantastic too. Especially on the all important memory front. 288GB is fantastic.
Both companies will leapfrog each other with new releases. Anyone who believes that there should only be a single vendor for all AI compute will quickly find themselves on the wrong side of history
This reminds me of those "192GB is fantastic" people that bought maxed-out M2 Ultras for AI inference. It can be awesome, but you need a substantial amount of interconnect bandwidth and powerful enough local compute before it's competitive. In products where AI is an afterthought, you're fighting against much different constraints than just having a lot of high-bandwidth memory.
I've always rooted for Team Red when they made an effort to do things open-source and transparently. They're a good role-model for the rest of the industry, in a certain sense. But I have to make peace with the fact that client-side AI running on my AMD machines isn't happening. Meanwhile, I've been using CUDA, CUDNN, CUBLAS, DLSS, on my Nvidia machine for years. On Linux!
What the customer does not see is how AMD must spend 2 times more money to produce a chip that is competitive against architecture that is soon 2 years old.
Probably because they aren't widely available yet. It is also a dual card to get that much memory, which is still less than 192GB and far less than 288GB.
> What the customer does not see is how AMD must spend 8-10 times more money to produce a chip that is competitive against architecture that is soon 2 years old.
as sibling mentioned the 188GB is for a pair. The memory bump is from enabling the 6th block of memory that is otherwise disabled on H100s. I assume an "NVL box" is still 8 total GPUs, so more like
Stocks of companies that develop extremely niche and technical things is a tiny sliver of the stock market that I actually think communities like HN would be better at valuing than the market.
Technology stocks are the only ones I personally day trade for that reason. Example: at the beginning of a pandemic lockdowns, any HN user could have anticipated increased internet usage and buy Cloudflare/Fastly stock and made a lot of money before the rest of the market realized that CDN companies will significantly benefit from that specific macro event.
I'm not convinced the market (or market analysts) have a deep understanding of Nividia's long-term advantage. If they did, we would have seen a much slower and steadier valuation increase rather than the meteoric rise. Meteoric stock price rise/fall = the market is having trouble valuing the stock.
In other words, stock prices don't add much to the conversation.
In my view, LLM is just the first step in the AI journey. The LLM boom will help NVidia to grow very fast and increase R&D. During this time, I expect new AI leaps that are not LLM-related. To be clear: I'm not talking about AGI, but rather, other practical advances.
AMD has been working on GPGPU at least as long as nVidia.
AMDs "CTM" SDK was released in 2006, same year as CUDA. In 2007 they released Stream SDK. Then they had "APP SDK" for a while, which iirc coincided with their opencl phase. And now they landed on rocm.
Meanwhile nvidia has kept trucking with just CUDA.
