Hardware Corner Dec 2025 ranking: RTX 3090, llama.cpp CUDA, Qwen3 8B Q4_K_XL GGUF, decode 46.59 t/s @65K ctx
46.59 tok/s per-request decode (batch 1, ~65536-token context)
Fields
id
benchmarks/hc-rank-rtx3090-qwen3-8b-q4kxl-tg-65kMetric
decode_tok_s
Value
46.59
Unit
tok/s per-request decode (batch 1, ~65536-token context)
Model
Qwen3 8B Q4_K_XL GGUF
Methodology
THIRD-PARTY LAB MEASURED (hardware-corner.net The Definitive GPU Ranking for LLMs, Allan Witt, updated December 9, 2025; the page states every number was captured on-hardware in their lab). llama.cpp llama-bench, Ubuntu 24.04, CUDA 12.8, 4-bit Q4_K_XL GGUF for every model (gpt-oss ships MXFP4). Token Generation table, 65K context column. ONE request, batch size 1: the table prints decode tok/s of a single stream at that KV depth, NOT aggregate across concurrent requests. The page does not publish the llama.cpp build hash or the exact prompt and output token counts beyond the context-depth bucket, and it does not say how many repetitions were averaged.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
format_id is the nearest standard GGUF in this repo; the page labels the quant Q4_K_XL (gpt-oss is MXFP4). A dash in the source table means the context did not fit and is not recorded. The RTX 4090 48GB modded row on the same page has no accelerator record and is excluded. Cells whose numeric value already existed in another record are skipped so the same measurement is not stored twice.
Notes
format_id is the nearest standard GGUF in this repo; the page labels the quant Q4_K_XL (gpt-oss is MXFP4). A dash in the source table means the context did not fit and is not recorded. The RTX 4090 48GB modded row on the same page has no accelerator record and is excluded. Cells whose numeric value already existed in another record are skipped so the same measurement is not stored twice.