Tesla P40 Llama-3.1-8B Q4_K_M decode (modelled)
35.0 tok/s decode (modelled)
Fields
id
benchmarks/comm-p40-llama31-8b-q4-tg-modelledMetric
decode_tok_s
Value
35
Unit
tok/s decode (modelled)
Model
Llama-3.1-8B
Methodology
vramguide.com modelled estimate from 347 GB/s GDDR5; prompt eval ~4.7 t/s. NOT a measured run.
Measured by
vendor
Thermal state
unknown
notes
MODELLED not measured. 347 GB/s is the binding constraint - interactive chat fine, long-context ingestion painful. No FP16 tensor path; GGUF Q4/Q8 DP4A only.
Notes
MODELLED not measured. 347 GB/s is the binding constraint - interactive chat fine, long-context ingestion painful. No FP16 tensor path; GGUF Q4/Q8 DP4A only.