llama.cpp b7083 Vulkan: gpt-oss-20b-Q8_0 decode 128 tok on amd-radeon-rx-9070
170.89 tok/s (single-request decode, 128 generated tokens, batch 1)
Fields
id
benchmarks/phx-vk-b7083-9070-gptoss20b-tg128Metric
decode_tok_s
Value
170.89
Unit
tok/s (single-request decode, 128 generated tokens, batch 1)
Model
gpt-oss-20b-Q8_0
Methodology
Phoronix Test Suite run of llama.cpp b7083 with the Vulkan backend, one shared host (Michael Larabel, 2025-12-08, result 2512070-NE-LLAMACPPV90). Intel Arc B580 on Mesa 26.0-dev ANV, AMD Radeon cards on Mesa 26.0-dev RADV, both under Linux 6.18. Geometric mean of N=3 runs. ONE request: the prompt-processing test prefills a single prompt and the text-generation test decodes a single 128-token completion. NOT an aggregate across concurrent requests. Batch size is the Phoronix llama.cpp default of 1 for these two tests. Quantization is Q8_0 as named in the result title. Power, clocks and thermals were not published. Card: amd-radeon-rx-9070. Model: gpt-oss-20b-Q8_0. DECODE of one request, text-generation-128 test, not aggregate.
Measured by
third_party
Reproducible
yes
notes
gpt-oss-20b-Q8_0 chart omits the RTX 5070 (it does not finish the run on that card in this result); the three cards filed here all completed. MoE decode, so the tok/s is not comparable to the dense 7B/8B rows above.
Notes
gpt-oss-20b-Q8_0 chart omits the RTX 5070 (it does not finish the run on that card in this result); the three cards filed here all completed. MoE decode, so the tok/s is not comparable to the dense 7B/8B rows above.