| 15 Mar 2026 |
 hexa | and nothing you could properly target as a baseline | 01:10:13 |
 dramforever | ... what? only ubuntu doesn't | 01:10:13 |
| hexa | fair, but as a distro we are much more rebuild heavy than all other distros | 01:11:50 |
| hexa | maybe except gentoo | 01:11:54 |
| dramforever | rva23 does not mean better performance | 01:12:04 |
| hexa | and what I've seen is that they barely beat an a55/a73 setup | 01:12:08 |
| hexa | the scaleway offer is for c910 cpus from what I can tell | 01:12:30 |
| hexa | will they even keep up with unstable-small rebuilds between staging merges? | 01:13:38 |
| dramforever | it does not mean we should turn on -march=rva23u64 while building | 01:14:01 |
| hexa | fair, whatever | 01:14:15 |
| hexa | I'm not insisting on that | 01:14:20 |
| dramforever | we can only do it with faster cpus | 01:14:34 |
| dramforever | if that's what you mean | 01:14:39 |
| hexa | yes | 01:14:43 |
| dramforever | i'm fine with that | 01:14:51 |
| hexa | I think we started bootstrapping aarch64-linux around some x-gene cpu from packet.net | 01:15:41 |
| hexa | not sure we had armv8 earlier than that | 01:15:53 |
| hexa | the community did provided an armv7 cache | 01:16:19 |
| hexa | * the community did provide an armv7 cache | 01:16:21 |
| dramforever | for reference, not the most representative benchmark, but from the people who had spacemit k3 ssh access, they got 15k on coremark with -march=rv64gc and 10k with v enabled | 01:17:20 |
| dramforever | yes coremark is more is better | 01:17:28 |
| dramforever | i think it's possible that "target rva23" becomes reasonable before nixos gets riscv64-linux in the upper tiers but it's not a reasonable target rn | 01:18:44 |
| hexa | nixos tests would be a big thing for baseline validation | 01:19:26 |
| hexa | maybe if the stars align we can get some tests onto nspawn | 01:20:48 |
| dramforever | yes, i think it is reasonable to say "we want fast chips with virtualization and X features and Y features for builders" | 01:21:04 |
| hexa | it was my impression that hardware virtualisation was not a consistent thing before the rv23a profile | 01:22:09 |
| hexa | * it was my impression that hardware virtualisation instructions were not consistenyl available before the rv23a profile | 01:22:33 |
| hexa | * it was my impression that hardware virtualisation instructions were not consistently available before the rv23a profile | 01:22:42 |
| dramforever | generally rva23 riscv machines are not available | 01:24:03 |
| dramforever | * | 01:24:11 |
