Chips news: GPU and server CPU in China. Qualcomm going to Samsung maybe, Microsoft Accelerator delayed
Lisuan new GPU is a first step, but performance far behind AMD/Nvidia. Loongson new server CPU look good, close to Intel chip of 2021. Major bottleneck remains SMIC capacity at 7nm.
Qualcomm looks close to qualifying Samsung’s 2nm to manufacture Snapdragon 8 Elite 2. This would imply a major yield breakthru for Samsung.
Microsoft ARM CPU Maia is in production, but its AI accelerator Braga seems delayed significantly – either performance or an architecture rethink.
China’s latest GPU
There are ~5 GPU start ups in China. Moore Threads, Biren and Lisuan Tech are probably the most advanced. Lisuan announced a GPU made on 6nm by SMIC.
Lisuan “claims that its performance can catch up with NVIDIA's RTX 4060, but the test results show that it is only at the level of the GTX 660 Ti from 13 years ago”.
I don’t know if that’s true. You’ll see report of tests on Geekbench here.
Conclusions:
1. China’s efforts to develop home-grown GPUs are accelerating
2. The Chinese firms have excellent engineer resources but they can’t catch up to Nvidia in a couple of years. It’s more a very long-term effort.
3. The bottleneck will remain manufacturing, ie SMIC ability to manufacture 6-7nm with high yields and the very hard task to made 5nm with DUV lithography machines.
Loongson has been designing CPU for 15 years probably.
Loongson tried many architectures, ARM, MIPS, and finally its own architecture (its own ISA LoongArchTM). Loongson chips are real, used mostly in Chinese government computers and servers, running Linux based OS and apps.
Loongson previous server CPU was the 3A5000 made by SMIC at 14-nm. You can buy a motherboard on Alibaba to try it out.
The new latest server CPU is called 3C6000, you’ll find details here and here. Performance seem close to Intel Xeon 8380 which was launched in 2021. This chip is made by SMIC at 7nm.
Conclusions:
1. China has had its own CPU for many years, performances are slowly catching up to AMD / Intel. Reaching performance similar to Xeon 8380 is a pretty big achievement.
2. The bottleneck is again SMIC capacity / yields at 7nm.
SMIC calls its 7-6nm process “N+2” and its 5nm process “N+3”
N+2 is clearly in production.
SMIC started N+2 production in 2022-23, compared to TSMC starting 7nm in 2028, EUV 7nm in 2020.
SMIC capacity about 20-30k wafer/month which is very small. For comparison TSMC has 330k wafer/month at 7-5-3nm.
Yields have improved considerably, reaching ~70% now which are commercial grade levels.
N+3 has been rumored many times, but there’s for now no chip verifiably made at N+3 by SMIC. Huawi Kirin X90 is supposed to be the first N=3 chip, TechInsight has a die analysis behind a paywall, leaks report that its actually still N+2.
Samsung 2nm good enough for Qualcomm?
This is from Korea’s Business Post.
Qualcomm is qualifying Samsung’s 2nm process for its upcoming Snapdragon 8 Elite 2 mobile processor. The chip has 2 variants. The base version will be made by TSMC at 3nm, the higher-end “S” version is undergoing validation on Samsung’s 2nm node.
This suggests that Samsung has had a yield breakthru at 2nm. Previous reports this year mentioned 30-40%. Commercial production for Qualcomm implies 60-70% within reach.
That would be excellent news for Samsung.
Microsoft AI chip delayed again?
That’s from The Information but there’s a paywall.
Microsoft has an ARM-based CPU in production, just starting, called MIAI and designed by Broadcom. This chip encountered its glitches and several versions were designed, Maia 100, 200, 300. Design started in 2019, launch was expected for 2024, now 2025.
Microsoft and Broadcom have also been working on AI Accelerators. This chip is called “Braga” and is “delayed by at least six months” with either/or 2 problems:
performance not good enough to justify production, meaning its more cost effective for Microsoft to buy Nvidia GPU
or Microsoft has changed its mind – it looks like Braga was originally a video accelerator, but Microsoft now wants something else.
I don’t know better.
