Builds / NVIDIA GeForce RTX 3090 24 GB
RTX 3090 + Ryzen 5 5600X + 32 GB DDR4: gpt-oss 20B at 211 tok/s, Qwen3.6 35B-A3B at 172, even gpt-oss 120B at 18 via NVMe paging (llm-benchmark.de)
Last updated:
- GPU(s)
- NVIDIA GeForce RTX 3090 24 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 24
- CPU
- AMD Ryzen 5 5600X (6-core)
Amazon · eBay (used) · Newegg - Motherboard
- ASUS ROG STRIX B550-A GAMING
Amazon · eBay (used) · Newegg - RAM
- 32 GB class (31 GB usable as reported)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- llama.cpp server-cuda b11804, -np 12, -c 49152; Timebench 3 combined prefill+generation workload, 1 concurrent request
- Reported total cost
- €1,518 (≈$1,701)
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| gpt-oss 20B | Q4_K_M (bartowski GGUF, MXFP4 experts) | llama.cpp CUDA b11804 | 211.1 | 4,232 (prompt length not stated) | context: 4,096 tokens per slot (49,152 total, 12 slots)📏 Timebench 3 combined prefill+generation run, 10.6 s total, 1 concurrent request |
| Qwen3.6 35B-A3B | Q4_K_M | llama.cpp CUDA b11804 | 172.4 | 2,763 (prompt length not stated) | context: 4,096 tokens per slot (49,152 total, 12 slots)📏 Timebench 3 combined run, 13.3 s total, fully on GPU (-ngl 999) |
| Nemotron 3.5 Lightning 30B-A3B | Q4_K_M (23.7 GiB, partial CPU offload) | llama.cpp CUDA b11804 | 101.0 | 1,039 (prompt length not stated) | context: 49,152 total, 12 slots📏 Timebench 3 combined run, 24.4 s total; automatic layer fitting put part of the model on the CPU |
| gpt-oss 120B | Q4_K_M (58.5 GiB, larger than VRAM + RAM) | llama.cpp CUDA b11804, rest mmapped from NVMe page cache | 18.1 | 96 (prompt length not stated) | context: 49,152 total, 12 slots📏 Timebench 3 combined run, 154 s total; weights exceed VRAM+RAM so it pages from NVMe |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Measured 24 Oct 2026 by Mario Alka on llm-benchmark.de (runs 2c110a gpt-oss-20b, 4aa5d8 Qwen3.6-35B-A3B, 79907e Nemotron 3.5 Lightning, 5a7454 gpt-oss-120b). The cost is the site's stored market-price total (GPU 850, CPU 121, board 149, RAM 248, PSU 150 EUR), not a receipt. This is a 'RAM matters less than you think' data point: with only 32 GB of system RAM the A3B MoE models run fully in VRAM at 170-210 tok/s, and even a 58.5 GiB gpt-oss 120B still decodes at 18 tok/s, paging from NVMe. Prefill is the cost there (20 s to first token).
https://llm-benchmark.de/benchmark/run-20261024-090500-5a7454
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/rtx-3090-ryzen-5600x-32gb-llm-benchmark-de/)
HTML
<a href="https://getrigbase.com/builds/rtx-3090-ryzen-5600x-32gb-llm-benchmark-de/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-3090-ryzen-5600x-32gb-llm-benchmark-de.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can RTX 3090 run Nemotron 3.5 Lightning 30B-A3B? · Can RTX 3090 run Qwen3.6 35B-A3B? · Can RTX 3090 run gpt-oss 120B? · Can RTX 3090 run gpt-oss 20B?