More
Submit build

Builds / 2x AMD Radeon AI PRO R9700 32 GB

2x Radeon AI PRO R9700 32 GB, vLLM with custom RDNA4 FP8 kernels: Qwen3.6 27B FP8 at 26.5 tok/s single stream, 55.3 with MTP (k=3)

Last updated:

Source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
Legacy models only. Every model measured on this rig now has a newer successor: Qwen3.6 27B → Qwen3.8 27B, Qwen3.6 27B → Qwen3.8 27B. The hardware numbers are still a useful reference.
GPU(s)
2x AMD Radeon AI PRO R9700 32 GB
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB)
64
CPU
not stated
Motherboard
not stated
RAM
not stated
PSU
not stated
Other
ROCm 7.2, vLLM fork mininmaxim/vllm branch esdmax-r9700-fp8 (ESDMAX FP8 GEMM kernels for gfx1201), tensor parallel across both cards, max-model-len 262,144
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.6 27B legacyFP8 (Qwen/Qwen3.6-27B-FP8)vLLM fork esdmax-r9700-fp8, ROCm 7.2, TP=226.51,242 @ 570-token prompt
est. TTFT @8k ≈ 6.6 s (calculated from measured prefill)
context: single stream📏 Single-stream decode after the kernel work (stock vLLM on the same box: 6-7 tok/s, FP8 silently fell back to FP32); author's bandwidth roofline 43.2; 570-token prompt prefill 1,242 tok/s
Qwen3.6 27B legacyFP8 (Qwen/Qwen3.6-27B-FP8)vLLM fork esdmax-r9700-fp8, ROCm 7.2, TP=2 + MTP k=355.3not stated⚡ spec-decode (MTP)context: single stream📏 qwen3_5_mtp speculative config, 3 speculative tokens, mean 2.94 accepted per step (k=1 44.4, k=2 51.9, k=4 57.3); temperature 0; 188 tok/s total at 8 concurrent requests

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 write-up in a public vLLM fork (committed 9 Jul 2026). CPU, RAM and board are not stated. The author's own llama.cpp Q8 run on the same box: 16 tok/s. Compare with Puget Systems' 2x R9700 build (stock vLLM FP8 15.88, tuned + MTP 62.75) - different builds and prompts.

Current GPU street price (estimate, not the builder's cost)Estimated GPU cost for this build today: ≈€3,696
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.
Source: https://github.com/mininmaxim/vllm/blob/esdmax-r9700-fp8/summary.md

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/2x-radeon-ai-pro-r9700-qwen36-27b-fp8-custom-kernels-mtp-mininmaxim.svg)](https://getrigbase.com/builds/2x-radeon-ai-pro-r9700-qwen36-27b-fp8-custom-kernels-mtp-mininmaxim/)

HTML

<a href="https://getrigbase.com/builds/2x-radeon-ai-pro-r9700-qwen36-27b-fp8-custom-kernels-mtp-mininmaxim/" rel="noopener"><img src="https://getrigbase.com/badge/2x-radeon-ai-pro-r9700-qwen36-27b-fp8-custom-kernels-mtp-mininmaxim.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · 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 2x Radeon AI Pro R9700 run Qwen3.6 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).