Hardware Corner: RTX 3090 Ti, llama.cpp CUDA, Qwen3 32B Q4_K XL GGUF, decode 35.42 t/s @16k ctx
35.42 tok/s per-request decode (batch 1, ~16384-token context)
Fields
id
benchmarks/hwcorner-rtx-3090-ti-qwen3-32b-q4kxl-decode-16kMetric
decode_tok_s
Value
35.42
Unit
tok/s per-request decode (batch 1, ~16384-token context)
Model
Qwen3 32B Q4_K XL GGUF
Methodology
THIRD-PARTY MEASURED (hardware-corner.net local-LLM GPU ranking, accessed 2026-10-14). llama.cpp CUDA build, Qwen3 weights in 'Q4K XL' GGUF (their standard quant; nearest repo format gguf-q4-k-m), batch 1. Metrics are per-request decode tok/s (tg, output tokens/s for a single stream) and prompt-processing pp512 tok/s; both measured at ~16384-token context. NOT aggregate batch throughput. Decode 35.42 t/s at 16k context for RTX 3090 Ti.
Measured by
third_party
Reproducible
yes
notes
Single-card consumer/workstation row. Value is per-request decode for ONE request, directly comparable to other hwcorner-* rows at the same context; the same page publishes short-context numbers that are higher.
Notes
Single-card consumer/workstation row. Value is per-request decode for ONE request, directly comparable to other hwcorner-* rows at the same context; the same page publishes short-context numbers that are higher.