3 ms·

No way it's 6x. The C910 is very similar to an ARM A72. On my own primes benchmark (https://hoult.org/primes.txt https://hoult.org/primes.txt) a Pi 4 at 1.5 GH

No way it's 6x. The C910 is very similar to an ARM A72. On my own primes benchmark (https://hoult.org/primes.txt https://hoult.org/primes.txt) a Pi 4 at 1.5 GHz takes 11.2 to 12.6 seconds depending which ISA you use (12.1 for 64 bit) while a HiFive Unmatched at 1.5 GHz (same U74 cores as the VisionFive) takes 15.3 seconds. So that's only about 25% on IPC, at least for integer codes in cache. The OoO A72 and C910 will get a bit more than that if there are a lot of cache misses. Add in 2.5/1.5 in MHz and you get about 2x overall. The real fun with RISC-V chips will start with the Intel "Horse Creek" which should be around RK3588 performance (similar to early Core i7?). A Horse Creek chip was shown (non-running) at a conference a couple of weeks ago and will probably be demo'd on a board, running, within the next month or so, and available on the next SiFive dev board maybe Q2 2023.

FullyFunctional 4y ago
It appears I have mistaken C910 with a different core. I haven't found trustworthy benchmarks but T-head's own numbers of 6 DMIPS/MHz and 7 CM/MHz (Dhrystones and CoreMark - which leave much to be desired) compared to 4.27/3.32 DMIPS/MHz (??) and 5.75 CoreMark/MHz seems to agree with you. I have U74 Geekbench results [1] and would like to see the same from C910. With that in mind, I agree that $1500 is a bit much. [1] GB 4.5.0 40 (Crypt 7, Int 49, FP 27), but don't compare to Arm and x86 architectures as unlike them, the RISC-V code is missing a lot of optimization.
- snvzz 4y ago
  >U74 geekbench results What version of U74? It has changed over time. Not only it got faster (both per clock, and changes to allow for clock increases), but it also gained extensions like B at some point. This is particularly important with cryptography in mind; I would assume the crypt result is about that.

FullyFunctional 4y ago

It appears I have mistaken C910 with a different core. I haven't found trustworthy benchmarks but T-head's own numbers of 6 DMIPS/MHz and 7 CM/MHz (Dhrystones and CoreMark - which leave much to be desired) compared to 4.27/3.32 DMIPS/MHz (??) and 5.75 CoreMark/MHz seems to agree with you. I have U74 Geekbench results [1] and would like to see the same from C910. With that in mind, I agree that $1500 is a bit much. [1] GB 4.5.0 40 (Crypt 7, Int 49, FP 27), but don't compare to Arm and x86 architectures as unlike them, the RISC-V code is missing a lot of optimization.

snvzz 4y ago
>U74 geekbench results What version of U74? It has changed over time. Not only it got faster (both per clock, and changes to allow for clock increases), but it also gained extensions like B at some point. This is particularly important with cryptography in mind; I would assume the crypt result is about that.