Yes, there is no question NVIDIA wants to lock you into CUDA and their hardware. But also, they’ve consistently demonstrated the most openness when it comes to model training, datasets, and research; even before the LLM era (e.g. StyleGAN).
There’s also modelscope.cn (china’s huggingface) which is worth checking out. I would not be surprised if one day, we have to use China VPNs to download open weight models.
I'll expand on the economics somewhat. My intuition is that you can like a moat left of you in the supply chain, but dislike moats to the right of you. (I'm using left and right as I picture this horizontally drawn. It is usually called vertical integration by economists.) But free competition in your market is worst of all. Free competition right outside your moat is pretty sweet, and that's were parent is commenting on.
Nvidia probably likes that ASML is a monopolist (it's called a monopsony). The price is high and Nvidia can't scale as hard as they want to (more general: capital intensive market). This monopsony makes sure that other chip companies can't rapidly scale up and try to beat Nvidia. That there only ever was one other GPU firm (I'm prehistoric; once there were more) and they bungled it on software, is pretty sweet for Nvidia.
On the side of their customers. It would be best for Nvidia if there is free competition for the outputs generated from there GPUs. This maximizes consumer surplus and thus demand. Maximum demand for tokens, is maximal demand feeded in their monopoly. If there is a monopoly right from you, demand is curtailed, and your value is limited.
And that exactly is why you see interest from token generators for chips. Bridge that moat and gain a larger value surplus. Both NVidia and, say, China actively undercutting the token-supplier value chain is quite interesting to watch. It's like the Opium wars with us as somewhat happy customers.
In this same vein, why isn't ASML raising thousands of billions for building their own (subsidiary) chip foundries, while raising prices and starving the market (a little) for their machines.
> In this same vein, why isn't ASML raising thousands of billions for building their own (subsidiary) chip foundries, while raising prices and starving the market (a little) for their machines.
You're looking this purely through an economic lens, while in reality geopolitical factors play a huge role in what ASML can and cannot do. The US government would likely take an extremely dim view of any new external competitor popping up for their chip foundry industry (especially with all the new US plants being built or planned) and would lean heavily on their vassal/ally the Netherlands to prevent this. Unlike with China, the US has more leverage over the Netherlands[1]
The US security state and US tech giants are joined at the hip, as they have been since the beginning of Silicon Valley[1], right through the Snowden revelations through to the present day[2].
There are - see the RTX Pro 6000, which has 96 GB.
There are a few problems though, primarily, a GPU with lots of VRAM and very high bandwidth is inherently very expensive (on top of which there is also the CUDA premium); AI use cases are better served by SoCs with lower (but still high) bandwidth and more RAM.
This is not a GPU, this is a laptop SoC with CPU and integrated GPU, and knowing nvidia it will probably be even more closed than an intel CPU. You will own even less of your hardware
When last looked at it, NVIDIA was not supporting OpenCL beyond 1.0.
Also, when running OpenCL, NVIDIA hardware disables multiple DMA engines, and allows only one memory transfer at a time to prevent OpenCL running as fast as CUDA.
Did NVIDIA finally allow open source drivers to access all parts and features of the card to allow feature parity? Last time I checked they were considering a plan for planning a solution to that.
I don't think bits like the GSP firmware will ever be open-sourced, but the opportunity to write better OpenCL drivers has always existed. Some of Nvidia's other proprietary driver backends (eg. GBM, Vulkan) are also decently neglected, but mostly out of disuse rather than malice. I don't think any of these things mean you don't own the hardware.
I love your insights which enlighten my blind spots most of the time, but I didn't blame NVIDIA for OpenCL's failure. I just noted a company's choices when it comes to a competing set of libraries w.r.t. to their native ones.
Having said that, I'll try compiling a OpenCL 3.0 program in the cluster, so I can report whether NVIDIA runs this software, and if yes, how well.
Yeah, however if Intel and AMD actually delivered a working 2.0 with proper support for C++ and Fortran, maybe the OpenCL 3.0 back to 1.0 reboot would not have been needed.
Likewise SYSCL although built on top of OpenCL 3.0 primitives, is mostly Intel, which also owns CodePlay, the company that delivered the first working SYSCL compute experience, again neither AMD nor Intel (until it bought CodePlay).
While nvidia 's drivers are closed source there's enough interest in running LLM's that you are not locked in using nvidia. Strix Halo chips have been around before dgx spark came out and deliver very similar performance.
What is the support like for them? Can they be seriously considered as alternatives to Nvidia or AMD cards? I have been looking into buying a GPU, specifically the Radeon R9700 AI Pro but noticed the Intel cards too and was not sure what you make of them.
Because making new AI models is a research activity, so the group within Nvidia (or other companies) that does that activity is called a research lab, or just "lab" for short. I don't think anyone is saying that all of Nvidia is a lab (that's just shorthand I guess).
What amuses me is that “lab” is shorthand for “laboratory” and virtually no computer research happens inside a room people would typically call a lab.
It’s a little like the trend of calling developers “engineers”. There’s no actual engineering in the traditional sense but I’m sure developers think it sounds cool to call themselves that.
That’s not what I said and the snarky remark about my experience is unwarranted (I have more experience in “software engineering” than many on here have been alive)
My point was just that I just find it amusing that people call themselves engineers when the code they produce is so far removed from the level of rigour one would expect in literally any other engineering industry.
I say this as someone who also has family and friends who are actual engineers, if they built bridges and buildings to the same standards that many developers write code, then people would die.
This isn’t meant as a criticism of developers, by the way. Just an observation at the vast differences in the domains and thus the tolerance for errors in the process.
Likewise for labs. I’ve worked in AI startups and the science departments are not something one would think of when you say “laboratory”. I get why the term is used, but it’s still amusing.
I know language isn’t static. It’s something that evolves, like how a “computer” used to refer to a person rather than a thing, but that doesn’t stop me from being amused. But maybe the real issue here is I take myself less seriously than others so I have that capacity to be amused by the titles I’ve held?
Because they are mostly research departments (aka a lab) turned into a corp structure. It’s not a new phenomenon, just that until recently labs didn’t get $1T valuations so you see it more now
Well if you find that obvious (I also do) then why do you not find their motives in acquiring HF onvious or that this move would be bad for the ecosystem as a whole? Cause it's all kinda the same thing.
"The proprietary shovel seller has some excellent tutorials on how to dig gold. Nobody else has such good step by step guides. Therefore them buying a shovel-agnostic tutorials and techniques method (that has a lot of info on using other shovels effectively) is justified."
Their (not especially great, compared to Chinese ones) models being open weight doesn't even come close to outweigh the effect of CUDA & Co being proprietary and closed.
A few years ago everyone said _Open_AI is the the most open labs. How did that turn out? Lots of coy, deceiving actions till the whole company was turned into whatever rent seeking amoral borg adjacent shell of it's former past it is now.
CUDA is proprietary for a pretty understandable reason. There's no good way for Nvidia to standardize it.
They supported OpenCL when Khronos floated the idea of a GPGPU standard to manufacturers, but OEMs didn't want to design scalable hardware or sponsor the software.
They even share many of their pre-training and even post-training datasets for Nemotron on HuggingFace; for example: https://huggingface.co/datasets/nvidia/Nemotron-Post-Trainin...
Which other lab shares this?
Yes, there is no question NVIDIA wants to lock you into CUDA and their hardware. But also, they’ve consistently demonstrated the most openness when it comes to model training, datasets, and research; even before the LLM era (e.g. StyleGAN).
There’s also modelscope.cn (china’s huggingface) which is worth checking out. I would not be surprised if one day, we have to use China VPNs to download open weight models.