inference/research
verifiedmedium · 0.75updated 2026-10-14

Dual EPYC 9334 (2x32c) llama.cpp DeepSeek-8B Q4 decode

27.8 tok/s

Fields

id
benchmarks/comm-dual-epyc-9334-deepseek-8b-decode
Metric
decode_tok_s
Value
27.8
Unit
tok/s
Model
DeepSeek-8B Q4
Methodology
Dual-socket, leaseweb 2026 bench; NUMA domain split means ~1.3-1.6x single socket not 2x
Measured by
third_party
Thermal state
unknown

Cross-references