verifiedmedium · 0.70updated 2026-10-14

RTX 5090 llama.cpp: Qwen3 8B Q4_K_XL prefill at 32k ctx (Hardware Corner)

3687.76 tok/s (pp~1024-2048 at ctx depth)

Fields

benchmarks/hc-rtx5090-qwen3-8b-q4kxl-32k-pp1024

Metric

prefill_tok_s

Value

3,687.76

Unit

tok/s (pp~1024-2048 at ctx depth)

Model

Qwen3 8B

Methodology

Hardware Corner (Allan Witt) llama-bench build 466c1911 on RTX 5090 32GB GDDR7 (EPYC 7642 host, Ubuntu 24.04). Qwen3 8B Q4_K_XL, -fa 1 -ngl 99, -p 1024/2048 -n 128 at context depth 32k. Per-request batch-1, NOT aggregate.

Measured by

third_party

Reproducible

yes

Cross-references

RTX 5090 llama.cpp: Qwen3 8B Q4_K_XL prefill at 32k ctx (Hardware Corner)

Fields

Cross-references

Accelerators (1)

Engine (1)

Format (1)

Interconnect (1)