AMD Strix Halo (Ryzen AI Max+ 395) GPU LLM Performance Tests

[Djip]

Does the Desktop have ROCM support

ROCm 6.4.4 adds formal support for both Windows and Linux for Strix Halo. https://www.amd.com/en/blogs/2025/the-road-to-rocm-on-radeon-for-windows-and-linux.html

Llama.cpp/vLLM Toolboxes for LLM inference on Strix Halo

Thanks! and like kyuz0 notice I can confirme that 6.4.4 is much faster than actual 7.0.1 release (and in paire with TheRock.) model size params backend ngl n_ubatch fa mmap test t/s qwen3 8B BF16 15.26 GiB 8.19 B ROCm 7.0.1 999 4096 1 0 pp512 325.95 ± 0.22 qwen3 8B BF16 15.26 GiB 8.19 B ROCm 6.4.4 999 4096 1 0 pp512 1132.26 ± 2.42

model	size	params	backend	ngl	n_ubatch	fa	mmap	test	t/s
qwen3 8B BF16	15.26 GiB	8.19 B	ROCm 7.0.1	999	4096	1	0	pp512	325.95 ± 0.22
qwen3 8B BF16	15.26 GiB	8.19 B	ROCm 6.4.4	999	4096	1	0	pp512	1132.26 ± 2.42