Strix Halo 8060S llama.cpp b6702 ROCm: Qwen3-30B-A3B UD-Q4_K_XL pp512 40.35 tok/s (depth 130560)
40.35 tok/s per-request prefill (llama-bench pp512, batch 1, depth 130560)
Fields
id
benchmarks/shhl-8060s-qwen3-30b-a3b-udq4-hipblaslt-pp512-d130560Metric
prefill_tok_s
Value
40.35
Unit
tok/s per-request prefill (llama-bench pp512, batch 1, depth 130560)
Model
Qwen3-30B-A3B UD-Q4_K_XL GGUF
Methodology
THIRD-PARTY MEASURED (strixhalo-homelab.d7.wtf AI/llamacpp-performance, accessed 2026-10-16). llama.cpp b6702 llama-bench, command printed on the page: -fa 1 -r 1 --mmap 0 -m Qwen3-30B-A3B-UD-Q4_K_XL.gguf -d 130560. Backend: ROCm 7.10.0a20251011 with ROCBLAS_USE_HIPBLASLT=1. Model Qwen3-30B-A3B Unsloth UD-Q4_K_XL GGUF on a Strix Halo iGPU (Radeon 8060S, gfx1151, Framework Desktop / GMKtec EVO-X2 class unified-memory box). context depth -d 130560 (131072-class context, 512-token prompt measured at that depth), single run (-r 1). Per-request batch-1, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
UD-Q4_K_XL is Unsloth dynamic Q4_K with an importance matrix; mapped to gguf-q4-k-m because there is no separate UD quant id. hipBLASLt rows are the same ROCm build with ROCBLAS_USE_HIPBLASLT=1 and are nearly identical to plain ROCm at both depths; kept because the page publishes them as a separate column. Long-context decode collapses on ROCm (64.17 to 4.98 tok/s) and holds up better on RADV (61.13 to 12.36 tok/s).
Notes
UD-Q4_K_XL is Unsloth dynamic Q4_K with an importance matrix; mapped to gguf-q4-k-m because there is no separate UD quant id. hipBLASLt rows are the same ROCm build with ROCBLAS_USE_HIPBLASLT=1 and are nearly identical to plain ROCm at both depths; kept because the page publishes them as a separate column. Long-context decode collapses on ROCm (64.17 to 4.98 tok/s) and holds up better on RADV (61.13 to 12.36 tok/s).