inference/research
verifiedmedium · 0.75updated 2026-10-15

Dual EPYC 9334 (32c x2, Genoa) gpt-oss-20b Q4 decode

26.0 tok/s decode (dual-socket, NUMA-distributed)

Fields

id
benchmarks/comm-dual-epyc-9334-gptoss-20b-decode
Metric
decode_tok_s
Value
26
Unit
tok/s decode (dual-socket, NUMA-distributed)
Model
gpt-oss-20b Q4-class
Methodology
LeaseWeb 2026 dual-socket EPYC 9334 llama.cpp bench; NUMA split across 24 channels
Measured by
third_party
Thermal state
unknown
notes
Dual socket does NOT double decode - ~1.3-1.6x a single socket in practice. Sibling row: DeepSeek-8B 27.8 t/s (comm-dual-epyc-9334-deepseek-8b-decode).

Notes

Dual socket does NOT double decode - ~1.3-1.6x a single socket in practice. Sibling row: DeepSeek-8B 27.8 t/s (comm-dual-epyc-9334-deepseek-8b-decode).

Cross-references