Builds / AMD Radeon AI PRO R9700 32 GB (640 GB/s)
Radeon AI PRO R9700 + Ryzen 7 5700X3D: Qwen3.8 27B quant ladder IQ4_NL to Q6_K (30 to 23 tok/s), 43-47 tok/s with MTP (akougkas)
Last updated:
- GPU(s)
- AMD Radeon AI PRO R9700 32 GB (640 GB/s)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 32
- CPU
- AMD Ryzen 7 5700X3D (8-core)
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- not stated
- PSU
- not stated
- Other
- Ubuntu 24.04.4, ROCm 7.2.4 (HIP), llama.cpp b10687, flash attention, q8_0 KV for the bench
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 27B | IQ4_NL | llama.cpp HIP b10687 (llama-bench) | 29.8 | 1,287 @ 2,048-token prompt est. TTFT @8k ≈ 6.4 s (calculated from measured prefill) | context: pp2048 / tg512📏 llama-bench tg512, fa on, q8_0 KV |
| Qwen3.8 27B | UD-Q4_K_XL | llama.cpp HIP b10687 (llama-bench) | 26.7 | 1,151 @ 2,048-token prompt est. TTFT @8k ≈ 7.1 s (calculated from measured prefill) | context: pp2048 / tg512📏 llama-bench tg512, fa on, q8_0 KV |
| Qwen3.8 27B | UD-Q5_K_M | llama.cpp HIP b10687 (llama-bench) | 24.5 | 946 @ 2,048-token prompt est. TTFT @8k ≈ 8.7 s (calculated from measured prefill) | context: pp2048 / tg512📏 llama-bench tg512, fa on, q8_0 KV |
| Qwen3.8 27B | UD-Q6_K | llama.cpp HIP b10687 (llama-bench) | 23.2 | 821 @ 2,048-token prompt est. TTFT @8k ≈ 10.0 s (calculated from measured prefill) | context: pp2048 / tg512📏 llama-bench tg512, fa on, q8_0 KV |
| Qwen3.8 27B | IQ4_NL + MTP draft 2 | llama-server HIP, 4 slots, 131k unified context | 43.0 | 979 (prompt length not stated) | ⚡ spec-decode (MTP)context: 12k-token prompt📏 Lower of two measurements (46.2 / 43.0), 66% draft acceptance, thinking-mode sampler |
| Qwen3.8 27B | IQ4_NL, no speculation | llama-server HIP, 4 slots, 131k unified context | 27.1 | 1,031 (prompt length not stated) | context: 12k-token prompt📏 Server run, same setup as the MTP row |
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.
Anthony Kougkas, 8 Sep 2026. Each step up the quant ladder costs a few tok/s: decode tracks the weight size because a dense 27B is bandwidth-bound. MTP draft 2 was the sweet spot (draft 3/4 accepted fewer tokens). Through his production router (f16 KV, thinking off) a cold 12k prompt decoded at 47.5 tok/s with 892 tok/s prefill. Vulkan in a real server was slower (23.8 tok/s, MTP a net loss). Four cached streams only add up to about 53 tok/s, so speculation helps latency, not batch throughput. His advice: don't go below IQ4_NL for speed.
- 1× Radeon AI Pro R9700: ≈€1,848 each (≈$2,071), new, approx. price as of 2026-10-24, lowest listing on billiger.de (PowerColor Radeon AI PRO R9700 32 GB GDDR6)
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.
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/r9700-5700x3d-qwen38-27b-quant-ladder-akougkas/)
HTML
<a href="https://getrigbase.com/builds/r9700-5700x3d-qwen38-27b-quant-ladder-akougkas/" rel="noopener"><img src="https://getrigbase.com/badge/r9700-5700x3d-qwen38-27b-quant-ladder-akougkas.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Radeon AI Pro R9700 run Qwen3.8 27B?