Re:Addressing Criticism of RISC-V Microprocessors
by ribit on Monday March 21, 2022 @04:06AM
(#30751500)
ET-minion can complete with GPUs because it is essentially a GPU. Most GPUs are in-order RISC machines with very wide SIMD ALUs and huge register banks. Whether it runs RISC-V or some proprietary ISA doesn’t really matter. How you build things does. And anyway, RISC-V vector ISA seems to be closely inspired by GPU ISAs to begin with.
Regarding SIMD: maybe you are right. I don’t know. Performance of RISC-V vectors here is an unknown factor. If VSETVL is zero-cost and the CPU can rename vector partitions without added latency, sure. But I remain skeptical until proven otherwise. The very design of RISC-V vector stuff screams amortized processing of large data blocks, not flexible low-latency operations on small data blocks. And notably, I could not find basic SIMD shuffle/interleave operations in the spec. That’s perfectly fine for large data vectors where you can work with multiple instruction, data-parallel vector registers and blends, but it would kill performance for many algorithms that work on limited number of lanes.
Of course CPUs are good for vector processing compared to a general-purpose CPU. That was not the point at all. The point is that unlike older architectures such as Cray, they were not designed specifically for general-purpose vector processing but for graphics processing. That is why solutions such as SOC-1 built specifically for genera-purpose vector processing can compete with graphics cards made by giants like Nvidia.
The article is talking about adding vector processing both to RISC-V chips aimed at general purpose processing as well as to specialized RISC-V cores which are primarily designed for vector-processing. SOC-1 is an example of this. It has 4 general purpose RISC-V cores called ET-Maxion, while also having 1088 small ET-Minion cores made for vector processing. However these are still RISC-V cores, rather than some graphics card SM core.
I don't get your argument about SIMD being great for latency. RISC-V requires that vector registers are at minimum 128-bit you you can use RVV as a SIMD instruction-set with 128-bit registers if you want.