RTX 4090 gpt-oss-20B token-gen @128K context (Hardware Corner)
70.0 tok/s
Fields
id
benchmarks/comm-rtx-4090-gpt-oss-20b-tg-128kMetric
decode_tok_s
Value
70
Unit
tok/s
Model
gpt-oss-20B
Methodology
THIRD-PARTY MEASURED (Hardware Corner llama.cpp llama-bench table). Q4-class GGUF, CUDA, batch-1 decode at 128K context; 190 t/s @4K on the same card - long-context KV read dominates.
Measured by
third_party
Reproducible
yes
notes
24GB GDDR6X, 1,008 GB/s. The 20B MoE at 128K ctx fits where a dense 32B does not; context-length is the speed knob.
Notes
24GB GDDR6X, 1,008 GB/s. The 20B MoE at 128K ctx fits where a dense 32B does not; context-length is the speed knob.