Step 3.5-Flash / Vulkan RADV native benchmark

on AMD Strix Halo 128GB · llama.cpp / Vulkan RADV / build unknown

Oct 13, 2026

Throughput

32.0 t/s gen · 237.0 tokens/s prompt processing (prefill)Prompt processing measures input prefill speed; it does not tell you the time to first token.

Quant

IQ3_XS (GGUF)

Flash Attention

System RAM

128 GB

Summary

User reports Step3.5-Flash at 32 t/s generation and 237 t/s prompt processing on an AMD Strix Halo 128GB system with one Radeon 8060S gfx1151 integrated GPU. Setup is the RADV backend with IQ3_XS weights, full GPU offload, batch 2048, 32 threads and flash attention 1. The figures come from separate pp512 and tg128 phases, not simultaneous 512-in/128-out throughput; the IQ3_XS artifact is 76 GiB on disk.