That's true, but I mean it'd be a hundred times worse if Deepseek did this on non-Nvidia cards. Which seems like only a matter of time if we're going to keep doing this.
> Which seems like only a matter of time if we're going to keep doing this.
What's funny is that people have been saying this since OpenCL was announced but today we're actually in a worse spot than we were 10 years ago. China too - their inability to replicate EULV advancements has left their lithography in a terrible place.
There's a reason China is literally dependent on Nvidia for competitive hardware. It's their window into export-controlled TSMC wafers and complex streaming multiprocessor designs. Losing access to Nvidia hardware isn't an option for them (or the United States for that matter) which is why the US pushes so hard for export controls. There is no alternative.
Well I wasn't optimistic about OpenCL in the past, because nobody will bother with that when they can pay a little (or even a lot) more for Nvidia and use CUDA. Even though OpenCL might work in theory with whatever tools someone is using, it's unpopular and therefore less supported. But this time is different.
Is it? Time will tell, but it wasn't "different" even during the crypto craze when CUDA was literally printing money. We were promised world-changing ASICs just like with Cereberas and Groq, and ended up with nothing in the end.
When AI's popularity blows over (and it will, like crypto), will the market have responded fast enough? From where I'm standing, it looks like every major manufacturer (and even the Chinese market) is trying the ASIC route again. And having watched ASICs die a very painful and unsatisfying death in the mining pools, I'd like to avoid manufacturing purpose-made ICs that are obsolete within months. I'm simply not seeing the sort of large-scale strategy that threatens Nvidia's actual demand.
I don't have any hope for ASICs. GPUs are going to stay, but if enough big countries are only able to obtain new non-Nvidia GPUs, big corps and even governments will find it worthwhile to build up the ecosystem around OpenCL or something else. Before now, CUDA had unstoppable momentum.
Crypto mining was just about hash rates, so I don't think it really mattered whether you used CUDA or not. Nvidia cards were just faster usually. People did use AMD too, but that didn't really involve building up the OpenCL ecosystem, just making it run one particular algo. They do also use ASICs for BTC in particular, I don't think that died.
