iBlessi/rtx-5080-llm-throughput

RTX 5080 Local LLM Throughput: Measured, Not Estimated

Measured decode and prefill tokens/second, load times, and VRAM residency for local LLMs on a single retail NVIDIA GeForce RTX 5080 (16 GB). 34 rows from two dated captures:

• 2026-08-12, 16 rows: gpt-oss:20b (MXFP4), Qwen 2.5 14B and 7B (Q4_K_M) and Llama 3.2 3B (Q4_K_M).
• 2026-08-26, 18 rows: the coding models qwen2.5-coder 7B and 14B, qwen3-coder 30B and devstral 24B (all Q4_K_M), plus a gpt-oss:20b re-capture for continuity.

Contexts are 4k and 16k (devstral 4k only). Prompts run from about 470 to about 2,600 tokens. The method did not change between the captures: every row is llm-throughput-ollama-v1 on driver 610.88 and ollama 0.32.1.

Method

• Every figure is the median of 3 runs using ollama's native eval_count/eval_duration counters, never wall-clock division.
• Every repetition prepends a unique nonce so each run pays a full prefill. Without it, warm runs hit ollama's prompt cache and report ~60,000 tok/s cache-hit rates that measure nothing.
• Generation pinned: temperature 0, seed 42, 256-token target output.
• Model identity pinned by digest per row (a tag like qwen2.5:14b can silently move upstream).
• Bench environment recorded per row: RTX 5080, driver 610.88, ollama 0.32.1, Windows 11, Ryzen 7 7800X3D.

Headline results

Model	Quant	Capture	Decode tok/s
Llama 3.2 3B	Q4_K_M	2026-08-12	298–313
gpt-oss 20B	MXFP4	2026-08-12 and 2026-08-26	178–190
Qwen 2.5 7B	Q4_K_M	2026-08-12	177–180
qwen2.5-coder 7B	Q4_K_M	2026-08-26	171–175
Qwen 2.5 14B	Q4_K_M	2026-08-12	94–97
qwen2.5-coder 14B	Q4_K_M	2026-08-26	93–95
qwen3-coder 30B (MoE), partial offload	Q4_K_M	2026-08-26	52–62
devstral 24B, 4k context only	Q4_K_M	2026-08-26	35–36

The 20.9B MXFP4 model decodes 2x faster than the dense 14B at Q4_K_M and matches the dense 7B. For the models that fit in VRAM, decode holds flat from 4k to 16k context at these prompt sizes. The partially offloaded qwen3-coder drops from about 62 to about 52.

Notes that belong next to those rows:

• qwen3-coder:30b is the first partial-offload row. Its 19GB Q4_K_M exceeds the card, so ollama ran it 74% GPU-resident. Its decode rate depends on free system RAM: the same model on the same card decoded at 83 tok/s with ~20GB free and under 1 tok/s while the box was paging.
• devstral:24b appears at 4096 context only. At 8k context its 14GB file plus KV cache spilled and decode collapsed below 1 tok/s on this box. The 16k cells were not captured for it.
• The gpt-oss:20b re-capture on 2026-08-26 runs 3-5% under its 08-12 rows, on a card six hours into continuous bench work. Both captures are real, and both carry their dates.

Caveats

Single retail unit, single driver, single runtime: these are this card's numbers. The vram_resident_mib column is whole-GPU nvidia-smi at measurement time and can include a previously-benched model still resident (rows measured alone are the lower readings).

Canonical source

Method detail, change log, the coding-eval grid and living updates: https://techfuelhq.com/data/rtx-5080-llm-throughput/?utm_source=huggingface&utm_medium=referral&utm_campaign=dataset-cards-2026-09&utm_content=rtx-5080-llm-throughput

License

CC BY 4.0. Attribute as "TechFuelHQ RTX 5080 LLM Throughput Dataset", linking to https://techfuelhq.com/data/rtx-5080-llm-throughput/