Jaap Buurman
Jaap Buurman
A new 7900xtx is slightly more expensive, and performing slightly worse than a used 3090. With performance optimizations it would be competitive, but right now the 3090 has the advantage....
@sirus20x6 Let me know if/when you start working on this. Would love to help test. I will create some graphs later to showcase how the 7900xtx currently scales with batchsize....
@thamwangjun Since MI100 is CDNA based, will these improvements only target that architecture? Or are improvements for RDNA3 based cards (eg 7900xtx) to be expected as well?
It does massively improve the situation. I REALLY like the idea of using labels at the volume level. That way, I can configure a single backup service, massively cutting down...
Docker compose labels all names volumes automatically, so generating a list of all named volumes is as easy as running the following command: `docker volume ls --filter label=com.docker.compose.volume`
Alternatively, you can also find all anonymous volumes via this command: `docker volume ls --filter label=com.docker.volume.anonymous`
Important to note is that this happens with amdvlk, radv and the proprietary pro drivers, so it's not driver specific either.
@0cc4m here is the full log with the `--verbose` flag included: https://gist.github.com/Mushoz/2a4a29e091d2fc7f320ac659510730b4
> Can you try reducing the ctx size? At half (64k) context it crashes with a OOM error for an attempted allocation of over 8 GB. At a quarter (32k)...
With `-fa` I can now load the full context