DeepSeek launched a free, open-source large language model in late December, claiming it was developed in just two months at a cost of under $6 million.
The Chinese companies probably use SIMC over TSMC from now on. They were able to do low volume 7 nm last year. Also, Nvidia and “China” are not on the same spot on the tech s-curve. It will be much cheaper for China (and Intel/AMD) to catch up, than it will be for Nvidia to maintain the lead. Technological leaps and reverse engineering vs dimishing returns.
Also, expect that the Chinese government throws insane amounts of capital at this sector right now. So unless Stargate becomes a thing (though I believe the Chinese invest much much more), there will not be fair competition (as if that has ever been a thing anywhere anytime). China also have many more tools, like optional command economy. The US has nothing but printing money and manipulating oligarchs on a broken market.
I’m not sure about 80/10 exactly of course, but it is in that order of magnitude, if you’re willing to not run newest fancy stuff. I believe the MI300X goes for approx 1/2 of the H100 nowadays and is MUCH better on paper. We don’t know the real performance because of NDA (I believe). It used to be 1/4. If you look at VRAM per $, the ratio is about 1/10 for the 1/4 case. Of course, the price gap will shrink at the same rate as ROCm matures and customers feel its safe to use AMD hardware for training.
So, my bet is max 2 years for “China”. At least when it comes to high-end performance per dollar. Max 1 year for AMD and Intel (if Intel survive).
If we are talking the manufacturing side, rather than design/software i am very curious to see how SIMC develops. You are absolutely right that there is a big advantage for the second mover, since they can avoid dead ends and already know on an abstract level what is working. And diminishing returns also help make gaps be slightly less relevant.
However i think we can’t just apply the same timeline to them and say “they have 7nm now” and it took others x years to progress from there to 5nm or 3nm, because these steps include the major shift from DUV to EUV, which was in the making for a very long time. And that’s a whole different beast compared to DUV, where they are also probably still relying on ASML machines for the smallest nodes (although i think producing those domestically is much more feasible). Eventually they’ll get there, but i think this isn’t trivial and will take more than 2 years for sure.
On the design side vs Nvidia the hyperscalers like Alibaba/Tencent/Baidu or maybe even a smaller newcomer might be able to create something competitive for their specific usecases (like the Google TPUs). But Nvidia isn’t standing still either, so i think getting close to parity will be extremely hard there aswell.
Of course, the price gap will shrink at the same rate as ROCm matures and customers feel its safe to use AMD hardware for training.
Well to what degree ROCm matures and closes the gap is probably the question. Like i said, i agree that their hardware seems quite capable in many ways, although my knowledge here is quite limited. But AMD so far hasn’t really shown that they can compete with Nvidia on the software side.
As far as Intel goes, being slow in my reply helps my point. Just today Intel canceled their next-generation GPU Falcon Shore, making it an internal development step only. As much as i am rooting for them, it will need a major shift in culture and talent for them to right the ship. Gaudi 3 wasn’t successful (i think they didn’t even meet their target of $500mio sales) and now they probably don’t have any release in 2025, assuming Jaguar Lake is 2026 since Falcon Shore was slated for end of this year. In my books that is the definition of being behind more than 1 year, considering they are not even close to parity right now.
Thanks for high effort reply.
The Chinese companies probably use SIMC over TSMC from now on. They were able to do low volume 7 nm last year. Also, Nvidia and “China” are not on the same spot on the tech s-curve. It will be much cheaper for China (and Intel/AMD) to catch up, than it will be for Nvidia to maintain the lead. Technological leaps and reverse engineering vs dimishing returns.
Also, expect that the Chinese government throws insane amounts of capital at this sector right now. So unless Stargate becomes a thing (though I believe the Chinese invest much much more), there will not be fair competition (as if that has ever been a thing anywhere anytime). China also have many more tools, like optional command economy. The US has nothing but printing money and manipulating oligarchs on a broken market.
I’m not sure about 80/10 exactly of course, but it is in that order of magnitude, if you’re willing to not run newest fancy stuff. I believe the MI300X goes for approx 1/2 of the H100 nowadays and is MUCH better on paper. We don’t know the real performance because of NDA (I believe). It used to be 1/4. If you look at VRAM per $, the ratio is about 1/10 for the 1/4 case. Of course, the price gap will shrink at the same rate as ROCm matures and customers feel its safe to use AMD hardware for training.
So, my bet is max 2 years for “China”. At least when it comes to high-end performance per dollar. Max 1 year for AMD and Intel (if Intel survive).
If we are talking the manufacturing side, rather than design/software i am very curious to see how SIMC develops. You are absolutely right that there is a big advantage for the second mover, since they can avoid dead ends and already know on an abstract level what is working. And diminishing returns also help make gaps be slightly less relevant.
However i think we can’t just apply the same timeline to them and say “they have 7nm now” and it took others x years to progress from there to 5nm or 3nm, because these steps include the major shift from DUV to EUV, which was in the making for a very long time. And that’s a whole different beast compared to DUV, where they are also probably still relying on ASML machines for the smallest nodes (although i think producing those domestically is much more feasible). Eventually they’ll get there, but i think this isn’t trivial and will take more than 2 years for sure.
On the design side vs Nvidia the hyperscalers like Alibaba/Tencent/Baidu or maybe even a smaller newcomer might be able to create something competitive for their specific usecases (like the Google TPUs). But Nvidia isn’t standing still either, so i think getting close to parity will be extremely hard there aswell.
Well to what degree ROCm matures and closes the gap is probably the question. Like i said, i agree that their hardware seems quite capable in many ways, although my knowledge here is quite limited. But AMD so far hasn’t really shown that they can compete with Nvidia on the software side.
As far as Intel goes, being slow in my reply helps my point. Just today Intel canceled their next-generation GPU Falcon Shore, making it an internal development step only. As much as i am rooting for them, it will need a major shift in culture and talent for them to right the ship. Gaudi 3 wasn’t successful (i think they didn’t even meet their target of $500mio sales) and now they probably don’t have any release in 2025, assuming Jaguar Lake is 2026 since Falcon Shore was slated for end of this year. In my books that is the definition of being behind more than 1 year, considering they are not even close to parity right now.