I agree with you “I’m still fairly confident from a macro perspective that forcing Nvidia/CUDA scarcity will create innovation in the Chinese ecosystem”
I guess that’s true—though if I understand your position on Sharp China, it’s that the Chinese ecosystem is going to evolve their way out of Nvidia dependence regardless. Given that, we might as well restrict it to the extent we can now?
At least that’s what I took from listening, but interested if that’s misunderstood on my part.
My take is the level of restriction does make a significant difference in how fast indigenization happens. I also think it’s naive to believe the GPU monopoly will be eternal (it’s the same shortage dynamic)… but the difference can be next year or next decade depending on how hard we squeeze on the supply. Most Chinese companies aren’t that interested in helping be Huawei’s guinea pigs (whatever the government asks) but they will put that effort in, if given no other option.
the political directive has been set, doesnt matter what the us does now. from the april politburo study session: Xi Jinping emphasized that to seize the initiative and gain the advantage in the AI field, breakthroughs must be achieved in fundamental theories, methods, tools, and more. We must continuously strengthen basic research, concentrate resources to overcome challenges in core technologies such as high-end chips and foundational software, and build an independent, controllable, and collaboratively functioning AI foundational hardware and software system [集中力量攻克高端芯片、基础软件等核心技术,构建自主可控、协同运行的人工智能基础软硬件系统]. We should use AI to lead a paradigm shift in scientific research and accelerate technological innovations and breakthroughs across various fields. https://sinocism.com/p/april-politburo-study-session-on
Indeed, and I think investment and flows will go that direction… but the Chinese private sector is not monolithic. Here’s why I think it might take a while if Nvidia GPUs are plentiful: it took years (nearly 15 years now) of hammering out CUDA’s bugs and optimizing it for scientific computing. That’s what makes it so good now.
Scientists and researchers were VERY patient and collaborative with Nvidia because there were no other great choices (vs supercomputers which were a huge jump up in cost and inaccessibility).
It’s not going to take 15 years to copy/follow. But without those real world reps, data, and refinement, CANN is going to improve at a painfully slow pace. It’s AMD’a problem with ROCm still. It’s also similar to fabs—you can’t “theoretically” improve them past a certain point. You need to actually run the process, get the shitty yields, get the data, and iterate until it gets good, which is partly what drives the benefits from scale.
Chinese companies are not altruistic/cooperative/etc enough to just put in the pain if they have a great alternative (quite the opposite given the competitive dynamics). Even if they have public pressure to make a show of it. They’ve been doing “research systems” for years that top the supercomputer charts with Huawei that are never used in production systems. They have to really commit to using Huawei as their real production systems, not just “parallel” or backup.
Anyway—I don’t have a crystal ball either, but that’s the source of my viewpoint on the topic. Appreciate you engaging and helping elaborate on your view!
I agree with you “I’m still fairly confident from a macro perspective that forcing Nvidia/CUDA scarcity will create innovation in the Chinese ecosystem”
I guess that’s true—though if I understand your position on Sharp China, it’s that the Chinese ecosystem is going to evolve their way out of Nvidia dependence regardless. Given that, we might as well restrict it to the extent we can now?
At least that’s what I took from listening, but interested if that’s misunderstood on my part.
My take is the level of restriction does make a significant difference in how fast indigenization happens. I also think it’s naive to believe the GPU monopoly will be eternal (it’s the same shortage dynamic)… but the difference can be next year or next decade depending on how hard we squeeze on the supply. Most Chinese companies aren’t that interested in helping be Huawei’s guinea pigs (whatever the government asks) but they will put that effort in, if given no other option.
the political directive has been set, doesnt matter what the us does now. from the april politburo study session: Xi Jinping emphasized that to seize the initiative and gain the advantage in the AI field, breakthroughs must be achieved in fundamental theories, methods, tools, and more. We must continuously strengthen basic research, concentrate resources to overcome challenges in core technologies such as high-end chips and foundational software, and build an independent, controllable, and collaboratively functioning AI foundational hardware and software system [集中力量攻克高端芯片、基础软件等核心技术,构建自主可控、协同运行的人工智能基础软硬件系统]. We should use AI to lead a paradigm shift in scientific research and accelerate technological innovations and breakthroughs across various fields. https://sinocism.com/p/april-politburo-study-session-on
Indeed, and I think investment and flows will go that direction… but the Chinese private sector is not monolithic. Here’s why I think it might take a while if Nvidia GPUs are plentiful: it took years (nearly 15 years now) of hammering out CUDA’s bugs and optimizing it for scientific computing. That’s what makes it so good now.
Scientists and researchers were VERY patient and collaborative with Nvidia because there were no other great choices (vs supercomputers which were a huge jump up in cost and inaccessibility).
It’s not going to take 15 years to copy/follow. But without those real world reps, data, and refinement, CANN is going to improve at a painfully slow pace. It’s AMD’a problem with ROCm still. It’s also similar to fabs—you can’t “theoretically” improve them past a certain point. You need to actually run the process, get the shitty yields, get the data, and iterate until it gets good, which is partly what drives the benefits from scale.
Chinese companies are not altruistic/cooperative/etc enough to just put in the pain if they have a great alternative (quite the opposite given the competitive dynamics). Even if they have public pressure to make a show of it. They’ve been doing “research systems” for years that top the supercomputer charts with Huawei that are never used in production systems. They have to really commit to using Huawei as their real production systems, not just “parallel” or backup.
Anyway—I don’t have a crystal ball either, but that’s the source of my viewpoint on the topic. Appreciate you engaging and helping elaborate on your view!
All true, but the companies are not really the deciders when the Party sets a priority.