Did a little research on RISC-V progress recently, so here it goes. It should be obvious by now that China has to move toward RISC-V and that it is moving in that direction. Basic things like 32-bit MCUs and low power CPUs and industrial control chips are the first ones they move to using RISC-V designs.
RISC-V is really good for these applications because they generally consume less power and can do more tasks than ARM chips at the same mature process node. So, I see Huahong building a bunch of 32-bit MCUs for AptChip and they are perfect for smart cars and machineries and such. And they can be produced with 55nm process. It's pretty good.
But then when we get into desktop, laptop and data center chips, you get areas that are dominated by x86 with some ARM processors. Smartphone CPUs and other data center chips are ARM processors. Even though RISC-V is more efficient, we haven't got to the point where they can offer strong enough performance to just match the processing power of x86/ARM CPUs.
Recently, we have seen a couple of RISC-V cores, including from SiFive with P670 where they reached older ARM CPU core. Remember this post?
At the time, we were very close to seeing 2nd generation XiangShan CPU getting taped out. Now, it looks like it will be taped out in Q1 and be shipped to customers in Q2/Q3.
You may wonder why 2nd gen Xiangshan is aimed at ARM A76 with just Spec2006 score 10/GHz, when Ventana is aiming for HPC and SiFive is aiming for ARM A78? Well, the simplest answer is that this RISC-V project has to be designed with a process that can be produced with domestic fabs. Not a lot of point designing for 4/5nm, when SMIC can't produce that for the next 2/3 years. Note, that both Veyron 1 and SiFive P670 use 5nm process. Now of course, most of P670 performance boost over Nanhu is that you can just clock 5nm chip a lot higher than 12/14nm chip. But even aside from that, I think you can get higher computation per GHz with a more advanced chip process. So, CAS is constrained by foundry unlike SiFive or Ventana.
So, I think it's quite impressive that they think they can achieve ARM Neoverse 2 level of computation with 3rd gen Xiangshan (Kunminghu) and that it will be ready sometimes in 2024 (so probably 6 months to 1 year later than Ventana) and that it will be able to replace ARM based on chips in data centers.
Now, remember that ARM Neoverse V2 was apparently off limit to Chinese chip designers based on the reporting from December.
As such, unless Chinese chip designers have special ISA/IP rights like Phytium and design their own special cores, they are pretty much handicapped if they continue to design ARM chips. That's why major players like Alibaba T-head, Huawei & Tencent are probably working with CAS on this through their joint effort in BOSC. They are making this open source so that others including all the major Chinese tech companies can help out with the effort and use them in their HPC chips. If Xiangshan does turn out to be comparable to ARM Neoverse V2 in performance, then they will be at a pretty good place.