Repository navigation
Replies: 1 comment 2 replies
|
@supomarlo This aligns with what I've observed. audio.cpp now has access to AMD DevLab, so I can run tests on real AI Max+ 395 Strix Halo Github runners. So far, HIP hasn't beaten Vulkan in general in my tests. There may be limitations in ggml's HIP backend on this hardware. Another possible reason is that some optimizations are enabled only on backends where they’ve been validated. HIP may therefore use a slower generic path, similar to the CPU backend, even when the hardware could support a faster one. Have you tried the new perf_mode=cuda_perf preset? It should also benefit HIP and Vulkan at the cost of higher VRAM usage, except for BreezeTTS2, where it's gated to CUDA, Vulkan, and Metal. |
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
All reactions