Builds / RTX 5090

RTX 5090 (450 W cap): stock Qwen3.8 27B all-NVFP4 on NInfer + DFlash2 at 165 tok/s prose, 326 code refactor, 400 module + tests; prefill 12,908 tok/s at 8K

Last updated: 11 Nov 2026

Source: sams.help ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source

GPU(s): RTX 5090
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB): 32
CPU: AMD Threadripper PRO (WRX80 workstation)
Amazon · eBay (used) · Newegg
Motherboard: WRX80 board
Amazon · eBay (used) · Newegg
RAM: about 123 GB
Amazon · eBay (used) · Newegg
PSU: not stated
Other: One RTX 5090 power-capped at 450 W, NVIDIA driver 580.178.04 (CUDA 13.0; images built on a CUDA 13.0.2 base), Docker, one container per card; ninfer-all fork of the NInfer C++/CUDA engine with the author's local (non-public) thinking patches; DFlash2 drafter bundled in the pack, 7 draft tokens, INT8 KV cache, --prefill-chunk 4096; 196,608-token context, 4 concurrent requests allowed (speeds are one request)
Reported total cost: not stated
Power: 450 W GPU cap

Measured speed (generation tokens/sec)

Model	Quant	Engine	tok/s	Prompt processing (tok/s)	Details
Qwen3.8 27B	all-NVFP4 NInfer v3 pack of the stock model (kaushikvira/Qwen3.8-27B-nvfp4full-dflash2-NInfer-v3)	NInfer (ninfer-all fork) + DFlash2 drafter, 7 draft tokens, INT8 KV	165	12,908 @ 8,321-token prompt est. TTFT @8k ≈ 0.63 s (calculated from measured prefill)	⚡ spec-decode (DFlash2)context: short prompt, 512 tokens out, thinking off📏 prose essay, temperature 0, cold prompt, median of 3 runs (162-167); measured 7 Nov 2026
Qwen3.8 27B	all-NVFP4 NInfer v3 pack of the stock model (kaushikvira/Qwen3.8-27B-nvfp4full-dflash2-NInfer-v3)	NInfer (ninfer-all fork) + DFlash2 drafter, 7 draft tokens, INT8 KV	326	7,856 @ 1,051-token prompt est. TTFT @8k ≈ 1.0 s (calculated from measured prefill)	⚡ spec-decode (DFlash2)context: short prompt, 512 tokens out, thinking off📏 code refactor (JavaScript to typed Python), temperature 0, median of 3 runs (307-335); measured 7 Nov 2026
Qwen3.8 27B	all-NVFP4 NInfer v3 pack of the stock model (kaushikvira/Qwen3.8-27B-nvfp4full-dflash2-NInfer-v3)	NInfer (ninfer-all fork) + DFlash2 drafter, 7 draft tokens, INT8 KV	400	11,181 @ 31,218-token prompt est. TTFT @8k ≈ 0.73 s (calculated from measured prefill)	⚡ spec-decode (DFlash2)context: short prompt, 512 tokens out, thinking off📏 code: a module plus its tests, temperature 0, median of 3 runs (381-405); measured 7 Nov 2026

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.

Owner-run recipe page (sams.help), 'What it runs today' section measured 7 Nov 2026 on the author's live server. Speculative decoding numbers: code is far more predictable than prose, so prose is the floor (about 2.5-2.8 drafted tokens kept per check vs about 5 for code). Same card, same author, stock NVFP4 export on vLLM with its built-in MTP-3 measured 111 tok/s on the author's thinking-on benchmark median (step 1, measured 28 Sep - 8 Oct). The page's headline 353 tok/s and the 635 tok/s tool-call figure use the Swift-1.5 fine-tune and/or unpublished local patches (per-request draft width, 4-bit drafter), so RigBase does not list them. In real agent use the author's server log shows about 285-291 tok/s.

Current GPU street price (estimate, not the builder's cost)

1× RTX 5090: ≈€7,140 each (≈$8,003), new, approx. price as of 2026-10-24, lowest listing on billiger.de (MSI GeForce RTX 5090 (lowest variant))

Estimated GPU cost for this build today: ≈€7,140
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.

Source: https://sams.help/recipes/qwen3-8-27b-5090-tok-s/

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

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

[![Measured on RigBase](https://getrigbase.com/badge/rtx-5090-threadripper-pro-qwen38-27b-nvfp4-ninfer-dflash2-sams-help.svg)](https://getrigbase.com/builds/rtx-5090-threadripper-pro-qwen38-27b-nvfp4-ninfer-dflash2-sams-help/)

HTML

<a href="https://getrigbase.com/builds/rtx-5090-threadripper-pro-qwen38-27b-nvfp4-ninfer-dflash2-sams-help/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-5090-threadripper-pro-qwen38-27b-nvfp4-ninfer-dflash2-sams-help.svg" alt="Measured on RigBase" height="20"/></a>

Preview: · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can RTX 5090 run Qwen3.8 27B?

Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).