Hacker News new | past | comments | ask | show | jobs | submit login

I did my own memcpy() benchmark on the first D1 board back in April 2021.

https://hoult.org/d1_memcpy.txt

I got about the same speed (1100 MB/s) for huge in-RAM copies. The in-cache speed peaked at 3771 MB/s for an 8 KB copy with custom code (shown on that page) using the D1's 128 bit vector registers, 2058 MB/s using the standard glibc function.

150 to 400 MB/s on the Pi Zero is very believable. The D1 is pretty strong RAM performance. The HiFive Unmatched was about 180 MB/s when I tested it. You need top-end DRAM controller IP and also a prefetch engine to get the good read speeds, and it's no surprise if an ARM11 doesn't have that. SiFive also doesn't have it on their self-produced chips. Hopefully Horse Creek has the good Intel IP for peripherals and RAM.

I have no idea what's going on with the Dhrystone. SiFive's in-order single-issue RISC-V cores get about 1.6 DMIPS/MHz on 32 bit cores, 1.7 on 64 bit cores. I haven't measured but I'd expect the C906 to also be around 1.7 i.e. 1700 VAX MIPS at 1 GHz.




This is great information, thank you! I hope SiFive's chips have improved too; I look forward to seeing the Horse Creek results.

1 GB/s or even 4 GB/s is still pretty anemic for single-computer laptops, but an order of magnitude improvement over Unmatched is great news.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: