Dual EPYC 9334 (2x32c) llama.cpp gpt-oss-20b Q4 decode
26.0 tok/s
Fields
id
benchmarks/comm-dual-epyc-9334-gpt-oss-20b-q4-decodeMetric
decode_tok_s
Value
26
Unit
tok/s
Model
gpt-oss-20b Q4
Methodology
THIRD-PARTY MEASURED (LeaseWeb 2026 EPYC-vs-L4 bench). Dual-socket Genoa, llama.cpp, ~26 t/s on gpt-oss-20b Q4. NUMA domain split means dual-socket gives ~1.3-1.6x a single socket, not 2x.
Measured by
third_party
notes
Sibling row comm-dual-epyc-9334-deepseek-8b-decode = 27.8 t/s on DeepSeek-8B Q4 in the same article.
Notes
Sibling row comm-dual-epyc-9334-deepseek-8b-decode = 27.8 t/s on DeepSeek-8B Q4 in the same article.