Hardware Corner — RTX 3080 Ti local LLM context-tier tables (Qwen3 Q4_K, March 2026)
benchmark
Fields
id
sources/hardware-corner-gpu-llm-bench-rtx-3080-tiPublisher
Hardware Corner (Allan Witt)
Kind
benchmark
Published
2026-03
Accessed
2026-10-16
notes
Hardware Corner per-GPU page, tables dated 'Updated Benchmarks: March 2026'. Prompt-processing and token-generation tok/s by context tier (4k/16k/32k/64k/128k) for Qwen3 8B Q4_K, Qwen3 14B Q4_K, and (where VRAM allows) gpt-oss 20B MXFP4. Single GPU, batch-1, per-request — NOT aggregate. Engine is the site's llama.cpp GGUF harness (same series as their llama-bench CUDA 12.8 / Q4_K_XL ranking); the per-GPU page does not reprint the llama-bench build hash, so confidence is below a build-pinned llama-bench log. Headline 'average' cards on some pages disagree with the table (e.g. 4070 Ti SUPER hero says 84.3 t/s while the 16k table says 47.2); ONLY the context-tier tables are recorded.
Notes
Hardware Corner per-GPU page, tables dated 'Updated Benchmarks: March 2026'. Prompt-processing and token-generation tok/s by context tier (4k/16k/32k/64k/128k) for Qwen3 8B Q4_K, Qwen3 14B Q4_K, and (where VRAM allows) gpt-oss 20B MXFP4. Single GPU, batch-1, per-request — NOT aggregate. Engine is the site's llama.cpp GGUF harness (same series as their llama-bench CUDA 12.8 / Q4_K_XL ranking); the per-GPU page does not reprint the llama-bench build hash, so confidence is below a build-pinned llama-bench log. Headline 'average' cards on some pages disagree with the table (e.g. 4070 Ti SUPER hero says 84.3 t/s while the 16k table says 47.2); ONLY the context-tier tables are recorded.
Cross-references
This record references no other records.