LeaseWeb: GPT-OSS-20B FP16 on dual EPYC 9334, 26.2 tok/s decode
26.2 tok/s (batch-1 decode, 24 threads)
Fields
id
benchmarks/leaseweb-epyc9334x2-gptoss20b-fp16-decodeMetric
decode_tok_s
Value
26.2
Unit
tok/s (batch-1 decode, 24 threads)
Model
gpt-oss-20b FP16
Methodology
THIRD-PARTY MEASURED (LeaseWeb blog, Fran Baigorri Hauen, 2026-04-05). HPE ProLiant DL385 Gen11, 2x EPYC 9334 (Zen4, 32c/64t each, 2.7GHz, 210W, 128MB L3/socket), 64GB DDR5 total (12-channel/socket). llama-bench (llama.cpp) local run + OpenLLM/llmperf API-level check. Test config: 512-token prompt, 128 tokens generated, 24 CPU threads for LLM rows. Values read off published charts/prose; CPU util 20-30%. FP16; 512+128 config; 24 threads; printed figure (MoE keeps FP16 viable).
Measured by
third_party
Reproducible
no
notes
MoE model - FP16 stays near Q4 speed because only ~3.6B params are active; the dense DeepSeek-8B is the contrast row.
Notes
MoE model - FP16 stays near Q4 speed because only ~3.6B params are active; the dense DeepSeek-8B is the contrast row.