It’s worth the entire read, even though it gets quite detailed.
A few takeaways:
• Huawei is a generation behind in chips, but its scale-up solution is arguably a generation ahead of Nvidia and AMD’s current products on the market.
• Huawei’s CloudMatrix 384 competes directly with Nvidia’s GB200 NVL72, and in some metrics is more advanced than Nvidia’s rack scale solution. The engineering advantage is at the system level, not just at the chip level, with innovation at the networking, optics, and software layers.
• Huawei’s Ascend chip is slower than Nvidia’s Blackwell, so Huawei package more of them into a server: “The tradeoff is simple: having five times as many Ascends more than offsets each GPU being only one-third the performance of an Nvidia Blackwell.”
• The biggest downside is power consumption, but that’s not nearly the problem in China that it is in the US. While the US has had no new nuclear power since the 1970s, China is rapidly expanding all sources of electrical power.
• The US export ban to China isn’t working:
- TSMC uses/licenses US Technology, and is not supposed to use that tech to make chips for China, but it is, despite being fined to the tune of $1B USD.
- HBM chips are banned to China, but systems that fall under the FLOPS compute limit that “happen” to use those chips are not banned. So, a Korea company (Faraday) builds systems using high-spec HBMs but low-spec compute and legally sells those to China. Faraday apparently even uses low-temperature solder in its systems so the Chinese have a easier time desoldering the chips to pair with Huawei’s Ascend chips.
It would seem that rather than slow China’s AI development prowess down, the main effect of the export ban is to further incentive China to develop their own tech.