LeaseWeb: DeepSeek-R1-8B FP16 on NVIDIA L4, 16.7 tok/s decode
16.7 tok/s (batch-1 decode)
Fields
id
benchmarks/leaseweb-l4-deepseek8b-fp16-decodeMetric
decode_tok_s
Value
16.7
Unit
tok/s (batch-1 decode)
Model
DeepSeek-R1-0528-Qwen3-8B FP16
Methodology
THIRD-PARTY MEASURED (LeaseWeb blog, same article/rig as the EPYC-9334 rows). NVIDIA L4 GPU, FP16 model, 512-token prompt + 128 generated. API-level OpenLLM/llmperf measurement.
Measured by
third_party
Reproducible
no
notes
Same-rig comparison row in the EPYC article; L4 ~2x the dual-socket CPU on FP16 dense.
Notes
Same-rig comparison row in the EPYC article; L4 ~2x the dual-socket CPU on FP16 dense.