Nvidia Announces H100 NVL – Max Memory Server Card for Large Language Models
anandtech.com
anandtech.com
The new "NVL" variant adds ~20% more memory per GPU by enabling the sixth HBM stack (previously only five out of six were used). Additionally, GPUs now come in pairs with 600GB/s bandwidth between the paired devices. However, the pair then uses PCIe as the sole interface to the rest of the system. This topology is an interesting hybrid of the previous DGX (put all GPUs onto a unified NVLink graph), and the more traditional PCIe accelerator cards (star topology of PCIe links, host CPU is the root node). Probably not an issue, I think PCIe 5.0 x16 is already fast enough to not bottleneck multi-GPU training too much.
I have seen some benchmarks from academia but nothing in the private sector.
I wonder if they thought they were moving too fast and wanted to milk amphere/ada as long as possible.
Not having any competition whatsoever means Nvidia can release what they like when they like.
AMD really needs to pick up the pace and make a solid competitive offering in deep learning. They’re slowly getting there but they are at least 2 generations out.
On desktops, only the 7000 series is kinda competitive for AI in particular, and you have to go out of your way to get it running quick in PyTorch. The 6000 and 5000 series just weren't designed for AI.
The existing ecosystems (cuda, pytorch etc) are all pretty garbage anyway -- aside from the massive number of tutorials it doesn't seem like it would actually be hard to build a vertically integrated competitor ecosystem ... it feels a little like the rise of rails to me -- is a million articles about how to build a blog engine really that deep a moat ..?
First of all you need hardware with cutting-edge chips. Chips which can only be supplied by TSMC and Samsung.
Then you need the software ranging all the way from the firmware and driver over something analogous to CUDA with libraries like cuDNN, cuBLAS and many others to integrations into pytorch and tensorflow.
And none of that will come for free, like it came to Nvidia. Nvidia built CUDA and people built their DL frameworks around it in the last decade, but nobody will invest their time into doing the same for a competitor, when they could just do their research on Nvidia hardware instead.
Realistically it's up to AMD or Intel.
https://www.cerebras.net/ Has innovative technology, has actual customers, and is gaining a foothold in software-system stacks by integrating their platform into the OpenXLA GPU compiler.
There's so little liquidity post-merge that it's only worth mining as a way to launder stolen electricity.
The bitcoin people still waste raw materials, and prices are relatively sticky with so few suppliers and a backlog of demand, but we've already seen prices drop heavily since then.
I got an email from vultr, saying that they're "officially taking reservations for the NVIDIA HGX H100", so I guess all public clouds are going to get those soon.
You can safely assume an entity bought as many as they could.
For anything that can be run remotely, it'll always be deployed and optimized server-side first. Higher utilization means more economy.
Then trickle down to local and end user devices if it makes sense.
[1] https://www.qualcomm.com/products/technology/processors/clou...
[2] https://github.com/quic/software-kit-for-qualcomm-cloud-ai-1...
If you want to build a business around LLMs, it makes a lot of sense to be able to run the core service of what you want to offer on your own infrastructure instead of rely on a 3rd party that most likely doesn't give more than 1% care about you.
The reasoning seems mostly obvious to me here: people do not care for the effort that decentralization requires. If given the option to run AI off some website to generate all you want, people will gladly do this over using their local hardware due to the setup required.
The unfortunate part is that it takes so much longer to create not for profit tooling that is just as easy to use, especially when the calling to turn that into your for profit business in such a lucrative field is so tempting. Just ask the people who have contributed to Blender for a decade now.
You can run many AI applications locally today that would have required a massive investment in hardware not all that long ago. It's just that the bleeding edge is still in that territory. One major optimization avenue is the improvement of the models themselves, they are large because they have large numbers of parameters, but the bulk of those parameters has little to no effect on the model output and there is active research on 'model compression', which has the potential to be able to extract the working bits from a model while discarding the non-working bits without affecting the output and realize massive gains in efficiency (both in power consumption as well as for running the model).
Have a look at the kind of progress that happened in the chess world with the initial huge ML powered engines that are beaten by the kind of program that you can run on your phone nowadays.
https://en.wikipedia.org/wiki/Stockfish_(chess)
I fully expect something similar to happen to language models.
Which is the point, decentralization will always be playing catch up here unless something really interesting happens. It has absolutely nothing to do with local compute power, that has always been on an incline. We just get fed scraps down the line.
If anything the steps are now so closely following each other that we have far more trouble tracking the societal changes and dealing with them than that we have a problem with the lag between technological advancement and its eventual commoditization.
Centralization of compute has not always won (even if that compute is mostly controlled by a single company). The failure of cloud gaming vs consoles, and the success of Apple (which is very centralized but pushes a lot of ML compute out to the edge) for example.
It was a slap in the face when the 4090 had the same memory capacity as the 3090.
A6000 is 5000 dollars, ain't no hobbyist at home paying for that.
If you are a business user then you must pay Nvidia gargantuan amounts of money.
This is the outcome of a market leader with no real competition - you pay much more for lower power than the consumer GPUs and you are forced into ujsing their business GPUs through software license restrictions on the drivers.
According to this, the difference seems to be that Studio Drivers are older and better tested, nothing else.
https://nvidia.custhelp.com/app/answers/detail/a_id/4931/~/n...
What am I missing in my understanding of Studio Drivers?
""" How do Studio Drivers differ from Game Ready Drivers (GRD)?
In 2014, NVIDIA created the Game Ready Driver program to provide the best day-0 gaming experience. In order to accomplish this, the release cadence for Game Ready Drivers is driven by the release of major new game content giving our driver team as much time as possible to work on a given title. In similar fashion, NVIDIA now offers the Studio Driver program. Designed to provide the ultimate in functionality and stability for creative applications, Studio Drivers provide extensive testing against top creative applications and workflows for the best performance possible, and support any major creative app updates to ensure that you are ready to update any apps on Day 1. ""
Given the size of LLMs, this should be possible with just a little bit of extra VRAM.
For example on 24GB, Llama 30B runs only in 4bit mode and very slowly, but I can imagine a RLHF finetuned 30B or 65B version running in at least 8bit would be actually useful, and you could run it on your own computer easily.
why do you think adding vram, but not cores will make it run faster?..
In any case, you're right it might not be as significant, however, the quality of the output increases with 8/16bit, and running 65B is completely impossible on 24GB
If they had trouble selling stock we would see this niche market get catered to.
So maybe more accurate is $200k+ a year and $20-30k on a workstation.
I grew up on $20k a year, the numbers in tech. are baffling!
ATI seems to be holding the idiot ball.
Port stable diffusion and clip to their hardware. Train an upsized version sized for a 48GB card. Release a prosumer 48gb card... get huge uptake from artists and creators using the tech.
Whether or not there is real competition depends entirely on whether Intels Arc line of GPUs stays in the market.
AMD strangely has decided not to compete. Its newest GPU the 7900 XTX is an extremely powerful card, close to the top of the line Nvidia RTX 4090 in raster performance.
If AMD had introduced it with an aggressively low price then then they could have wedged Nvidia, which is determinbed to exploit it's market dominance by squeezing the maximum money out of buyers.
Instead, AMD has decided to simply follow Nvidia in squeezing for maximum prices, with AM prices slightly behind Nvidia.
It's a strange decision from AMD who is well behind in market and apparently seems disinterested in increasing that market share by competing aggressively.
So a third player is needed - Intel - it's alot harder for three companies to sit on outrageously high prices for years rather than compete with each other for market share.
Much less powerful GPUs represent better value but the market is ridiculously overpriced at the moment.
Since Intel GPUs are again TSMC manufactured, you really aren't going to see price improvements unless Intel subsidizes all of this.
This is not correct.
You are incorrect that this is the root cause of GPU prices being sky high.
If manufacturing cost was the root cause then it would be simply impossible to bring prices down without losing money.
The root cause of GPU prices being so high is lack of competition - AMD and Nvidia are choosing to maximise profit, and they are deliberately undersupplying the market to create scarcity and therefore prop up prices.
"AMD 'undershipping' chips to help prop prices up" https://www.pcgamer.com/amd-undershipping-chips-to-help-prop...
"AMD is ‘undershipping’ chips to balance CPU, GPU supply Less supply to balance out demand—and keep prices high." https://www.pcworld.com/article/1499957/amd-is-undershipping...
In summary, GPOU prices are ridiculously high because Nvidia and AMD are overpricing them because they believe this is what gamers will pay, NOT because manufacturing costs have forced prices to be high.
- The Intel Falcon Shores XPU is basically a big GPU that can use DDR5 DIMMS directly, hence it can fit absolutely enormous models into a single pool. But it has been delayed to 2025 :/
- AMD have not mentioned anything about the (not delayed) MI300 supporting DIMMs. If it doesn't, its capped to 128GB, and its being marketed as an HPC product like the MI200 anyway (which you basically cannot find on cloud services).
Nvidia also has some DDR5 grace CPUs, but the memory is embedded and I'm not sure how much of a GPU they have. Other startups (Tenstorrent, Cerebras, Graphcore and such) seemed to have underestimated the memory requirements of future models.
That's the problem. Good DDR5 RAM's memory speed is <100GB/s, while nvidia could has up to 2TB/s, and still the bottleneck lies on memory speed for most applications.
Anyway, what I was implying is that simply fitting a trillion parameter model into a single pool is probably more efficient than splitting it up over a power hungry interconnect. Bandwidth is much lower, but latency is also slower, you are shuffling much less data around.
I'm not saying they shouldn't bother with RAM at all, mind you. But given some target price, it's a balance thing between compute and RAM, and right now it seems that RAM is the bigger hurdle.
Are you sure about that?
Depending on the model the performance is sometimes not all that different. I believe for solely inference on some models the speed difference may barely be noticeable, where for other training activities it may make 10+% difference [1]
[0] https://pytorch.org/tutorials/intermediate/model_parallel_tu...
[1] https://huggingface.co/transformers/v4.9.2/performance.html
> “The reason we took [NVLink] off is that we need I/O for other things, so we’re using that area to cram in as many AI processors as possible,” Jen-Hsun Huang explained of the reason for axing NVLink.[0]
"NVLink is bad for your games and AI, trust me bro."
But then this card, actually aimed at ML applications, uses it.
0. https://www.techgoing.com/nvidia-rtx-4090-no-longer-supports...
It's also enormously more expensive and I'm not sure if you can buy it new without getting the nvidia compute server.
Previously, GPUs were designed for gamers, and no game really "needs" more than 16 GB of VRAM. I've seen reviews of the A100 and H100 cards saying that the 80GB is ample for even the most demanding usage.
Now? Suddenly GPUs with 1 TB of memory could be immediately used, at scale, by deep-pocket customers happy to throw their entire wallets at NVIDIA.
This new H100 NVL model is a Frankenstein's monster stitched together from whatever they had lying around. It's a desperate move to corner the market early as possible. It's just the beginning, a preview of the times to come.
There will be a new digital moat, a new capitalist's empire, built upon on the scarcity of cards "big enough" to run models that nobody but a handful of megacorps can afford to train.
In fact, it won't be enough to restrict access by making the models expensive to train. The real moat will be models too expensive to run. Users will have to sign up, get API keys, and stand in line.
"Safe use of AI" my ass. Safe profits, more like. Safe monopolies, safe from competition.