It’s kinda to be expected. US export controls are preventing the sale of NVida chips to Chinese companies. Those companies can either adapt or go out of business. So, they are adapting. China also have a sizable semi-conductor industrial base; so, we should expect to see some switch to domestic production to de-risk their supply chain. While it will be a long time (if ever) for that domestic production capability to catch up with US based companies, the current situation does give them an incentive and source of funding to try. It will be interesting to see, going forward, if Huawei can improve on their technology and processes to challenge NVidia.
It will be interesting to see, going forward, if Huawei can improve on their technology and processes
Why is this even a question? China is leading the world in high tech research.
to challenge NVidia
This is where the question is, but if we examine it, why would it be a question of if instead of when? What magic does NVidia have that the largest population in the world with the fastest growing economy in the history of the world can’t compete with?
The spun off company from chinas nvidia branch “Moore Threads” has a gpu out that you can buy, but drivers are extremely terrible on it.
Intel is an example of a major corporation who jumped into the dgpu game a few years ago, and they ran into several driver problems and still have many to this day.
The hardest part of gpu design nowadays isn’t the hardware part, but the software side
of course, its just that its a huge task to undertake. People still give AMD shit for their drivers and they’ve been doing it for over a decade. It’s a huge problem for any company to build a competent driver team, especially for graphics, be it on PC or on mobile (e.g Qualcomm has a negative stigma for not supporting its devices long). The when for it is not in any short time window and would be a task that would take several years realistically. One way to fast track it would be approaching in an open source way (e.g AMD/Intel linux GPU drivers) but at a corporate level, drivers for those kinds of gpus usually cost way more to produce (part of the reason why workstation and server gpus are magnitudes more expensive than their consumer counterpart)
The A100 is widely considered to be about half the performance of Nvidia’s newest H100. More importantly for LLMs, it’s limited to about half the memory of the H100.
Add comment