native PP512 and separate TG128

on AMD Radeon 8060S integrated GPU (Strix Halo; installed RAM unreported) · llama.cpp b5348 / 09232370 / ROCm rocWMMA

Oct 14, 2026

Throughput

4.7 t/s gen · 34.4 tokens/s prompt processing (prefill)Prompt processing measures input prefill speed; it does not tell you the time to first token.

Quant

i1-Q4_K_M (GGUF)

Flash Attention

Summary

User reports Shisa V2 Llama 3.3 at 4.7 t/s generation and 34.4 t/s prompt processing on a Ryzen AI Max+ 395 / Radeon 8060S under Linux. Setup is llama.cpp with the i1-Q4_K_M quant, HIP plus WMMA and FA, ngl 99, on one local ROCm device; the model is 39.59 GiB and 70.55B parameters. The 34.4 t/s is PP512 and the 4.7 t/s is a separate TG128 phase, not a same-request pair; context depth is unreported.