Builds / NVIDIA Tesla P40 24 GB
Tesla P40 24 GB (Pascal, 2016): Qwen3.6 35B-A3B UD-Q4_K_M at 48.99 tok/s, Qwen3.8 27B UD-Q4_K_M at 12.13 (stock llama.cpp, CUDA 12.8)
Last updated:
- GPU(s)
- NVIDIA Tesla P40 24 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 24
- CPU
- Intel Xeon E5 v3 (used workstation)
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- 15 GiB
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Pascal GP102, compute capability 6.1, no tensor cores, FP16 at 1/64 of FP32; measured 235.9 GB/s memory bandwidth; PCIe Gen3 x16; driver 580.178.04; CUDA 12.8; llama.cpp dd1ea52 built with CMAKE_CUDA_ARCHITECTURES including 61 (needed for DP4A); Ubuntu 24.04.4
- Reported total cost
- not stated
- Power
- 250 W limit (Dense run 80 C / 250 W, MoE run 70 C / 188 W)
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | Unsloth UD-Q4_K_M (20.60 GiB) | llama.cpp dd1ea52 (CUDA 12.8, sm_61) | 48.99 | 661 @ 512-token prompt est. TTFT @8k ≈ 12.4 s (calculated from measured prefill) | context: llama-bench tg128📏 llama-bench -p 512 -n 128 -r 5, one card; prompt pp2048 needed -ub 1024 to fit (951.39 tok/s), -ub 2048 ran out of memory |
| Qwen3.8 27B | Unsloth UD-Q4_K_M (15.32 GiB) | llama.cpp dd1ea52 (CUDA 12.8, sm_61) | 12.13 | 256 @ 512-token prompt est. TTFT @8k ≈ 32.0 s (calculated from measured prefill) | context: llama-bench tg128📏 llama-bench -p 512 -n 128 -r 5; 85% of the card's measured-bandwidth ceiling (14.3 tok/s); depth sweep with -fa 1 -ub 2048: tg32 12.09 at 0, 11.33 at 16K, 10.64 at 32K |
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.
Zenn article (published 3 Sep 2026, updated 16 Sep), measured on the author's own hardware with the same llama.cpp build, script and models across cards; the article states an LLM agent ran the setup and measurements under the author's direction. Same-test comparison cards: Tesla V100 PCIe 32GB 99.08 (MoE) / 34.61 (dense) tok/s, NVIDIA CMP 170HX (patched driver, 64 GB) 121.16 / 40.21, Tesla P100 16GB 12.90 dense (MoE does not fit). Takeaway from the source: on old low-bandwidth cards pick a MoE; the P40 ran the 35B-A3B MoE 4x faster than the dense 27B.
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/tesla-p40-24gb-pascal-qwen36-qwen38-zenn/)
HTML
<a href="https://getrigbase.com/builds/tesla-p40-24gb-pascal-qwen36-qwen38-zenn/" rel="noopener"><img src="https://getrigbase.com/badge/tesla-p40-24gb-pascal-qwen36-qwen38-zenn.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use