111–120 of 444 posts
Why, and what does it mean, for Nvidia to announce fiscal results a year ahead of time. Is it just promise to sell chips in advance, so that's how far it's booked, do they own a Time Machine...?
Financial years are named by the calendar year that they end in, so FY24 is the financial year ending in 2024.
Incredible company. It’s absolutely insane how far ahead they are with the investments they made over a decade ago. So nice to see a “hard” engineering (from silicon to software) SV-founded company getting all this recognition. Especially after what has felt like a decade of SV hype software companies dominating the mainstream financial markets pre-pandemic with a spate of overpriced IPOs or large ad-revenue generati…
The moniker of "hard" engineering is neither precise nor useful. What makes engineering hard? Is solving problems with distributed systems, even if these systems are for ads, hard? Or do you mean hardware? In that case even Nvidia is not hard enough since they don't fabricate their own chips. Or do you mean designing hardware? Then what makes writing system verilog at a desk hard but writing Python not hard?
Earlier quoted context omitted.
They’re all pretty motivated, they’ve been motivated for years, and almost nothing is happening. This situation isn’t exactly a poster child for the Efficient Markets Hypothesis. Every year just sounds like “Nvidia’s new consumer GPUs are adding new features, breaking previous performance ceilings, running games at huge resolutions and framerates. Their datacenter cards are completely sold out because they can spin s…
> This situation isn’t exactly a poster child for the Efficient Markets Hypothesis. I'm unsure why you're criticizing the Efficient Markets Hypothesis or even using it here, but you need to also analyze this with some time horizon because the market and marketplaces are not static.
Incredible company. It’s absolutely insane how far ahead they are with the investments they made over a decade ago. So nice to see a “hard” engineering (from silicon to software) SV-founded company getting all this recognition. Especially after what has felt like a decade of SV hype software companies dominating the mainstream financial markets pre-pandemic with a spate of overpriced IPOs or large ad-revenue generati…
Are they so far ahead? AMD GPUs get comparable results as of late on Stable Diffusion. Software and hardware from competitors will catch up, crunching 4/8/16 bit width numbers is no rocket science.
The recent Stable Diffusion results are great news, but also don't include comparisons to an Nvidia card using the same optimizations. Nvidia claims that Microsoft Olive doubles performance on their cards too, so it might be a bit of a wash: https://blogs.nvidia.com/blog/2023/05/23/microsoft-build-nvi...
Plus, none of those optimizations were any more open than CUDA (since it used DirectML).
> crunching 4/8/16 bit width numbers is no rocket science.
Of course not. That's why everyone did it: https://onnxruntime.ai/docs/execution-providers
The problem with that "15 competing standards" XKCD is that normally one big proprietary standard wins. Nvidia has the history, the stability, the multi-OS and multi-arch support. The industry can definitely overturn it, but they have to work together to obsolete it.
Nvidia's undervalued. Once enterprise adoption of AI picks up, demand for chips will increase 2-3 times further. I'm told Nvidia's building their own fab in Southeast Asia over the next few years. This will massively boost their output.
That and possibly way more than that is already priced in. Nvidia's stock is extremely expensive not because of they are making now (which is not a lot relative to valuation, they just barely surpassed Intel this quarter in revenue) but because investors expect pretty much exponential growth over the next few years..
Earlier quoted context omitted.
It remains debatable whether mass enterprise adoption of AI would happen first, or Nvidia's competitors coming up with equivalent chips would happen first.
On the surface, it's not debatable. Enterprises are going full steam ahead on AI. Building out an ecosystem to challenge Nvidia seems like a decade long battle, if it's even possible.
In any case Nvidia is expecting to ship ~550k H100s in 2023, hardly enough to satisfy every user.
Tesla decided to in-house. TPUv4 and Gaudi2 exceeded A100 performance, they just never hit scale or the market and then Hopper added optimization for transformers rendering these chips relatively obsolete.
Nvidia's lead is not unassailable and it seems incredibly unlikely that they would not face serious competition within the next 2-3 years given the $ being thrown around.
I've seen a regular stream of reports on HN about people "sort of" getting AI done on laptops and non and lowly GPU machines. Is it unreasonable or far-fetched to imagine that someone figures out how to efficiently get it all done without GPUs and pull the rug out from under Nvidia?
All you have to do is take 5 seconds in a typical code base to determine that the way we write software today isn't exactly... ideal. Given another 6-12 months, I cannot comprehend another ~OOM not being extracted somewhere simply by making the software better.
Earlier quoted context omitted.
> Can't they just hire all of nvidia's developers and pay them 5x as much? As time goes on I don’t see how you break the CUDA moat even if you had all of nvidia Al’s engineers. CUDA means you need everyone in AI to target your new (hopefully open) platform and that platform is faster than CUDA is. Given how most frameworks of the last 10 years have been optimized for CUDA you would need to turn around a global sized…
Well I'm not super experienced with GPU development but aren't most people using packages built on top of CUDA like pytroch etc? Would it be impossible to throw tons of resources at those packages so they handle whatever intel comes up with as well as they handle CUDA? If Intel is 10% slower but 50% cheaper and the open source stack you use has been heavily updated to work well with Intel drivers would that not be an…
Yes, and in fact both AMD and Intel have libraries. You can run Stable Diffusion and suchlike on AMD GPUs today, apparently. And you can export models from most ML frameworks to run in the browser, on phones and suchlike.
> If Intel is 10% slower but 50% cheaper [...] would that not be an enticing product?
Sometimes, yes. Some of the largest models apparently cost $600,000 in compute time to train [1], so halving that would be pretty appealing.
However, part of the reason for nvidia's dominance is that if you're hiring an ML engineer for $160,000/year spending $1,600 to give them an RTX 4090 is chump change.
[1] https://twitter.com/emostaque/status/1563870674111832066