verifiedmedium · 0.75updated 2026-10-15

EPYC 9554 (Genoa) DeepSeek-R1-70B Q4_K_M decode

7.11 tok/s decode, single-stream

Fields

benchmarks/comm-epyc-9554-deepseek-r1-70b-q4-decode

Metric

decode_tok_s

Value

7.11

Unit

tok/s decode, single-stream

Model

DeepSeek-R1-70B Q4_K_M

Methodology

llama.cpp llama-bench on EPYC 9554, 12x DDR5-5600, --numa distribute

Measured by

third_party

Thermal state

unknown

notes

Same host as the Llama-3.3-70B 7.12 t/s row; near-identical because decode is bandwidth-bound.

Same host as the Llama-3.3-70B 7.12 t/s row; near-identical because decode is bandwidth-bound.