Builds / AMD Radeon 8060S (Strix Halo, unified)
Framework Desktop (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B UD-Q4_K_M at 12.82 tok/s, Gemma 4 31B Q4_K_M at 11.27 (llama-bench tg128)
Last updated:
- GPU(s)
- AMD Radeon 8060S (Strix Halo, unified)
- VRAM / unified memory (GB)
- 128
- CPU
- AMD Ryzen AI Max+ 395
Amazon · eBay (used) · Newegg - Motherboard
- Framework Desktop
Amazon · eBay (used) · Newegg - RAM
- 128 GB shared memory
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- llama.cpp commit f5222f25c built in Release mode, Vulkan with KHR cooperative matrices; Mesa RADV 26.2.5-arch1.1; the driver exposes 111.56 GiB to Vulkan with about 109.84 GiB free at the start; models run sequentially
- 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 | UD-Q4_K_M (15.33 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 12.82 | 414 @ 512-token prompt est. TTFT @8k ≈ 19.8 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
| Qwen3.8 27B | UD-Q6_K (20.47 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 9.71 | 401 @ 512-token prompt est. TTFT @8k ≈ 20.4 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
| Qwen3.8 27B | Q8_0 (27.05 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 7.5 | 477 @ 512-token prompt est. TTFT @8k ≈ 17.2 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
| Gemma 4 31B | Q4_K_M (17.07 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 11.27 | 360 @ 512-token prompt est. TTFT @8k ≈ 22.8 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
| Gemma 4 31B | Q6_K (23.47 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 8.1 | 348 @ 512-token prompt est. TTFT @8k ≈ 23.5 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
| Gemma 4 31B | Q8_0 (30.39 GiB) | llama.cpp (Vulkan, commit f5222f25c) | 6.46 | 412 @ 512-token prompt est. TTFT @8k ≈ 19.9 s (calculated from measured prefill) | context: llama-bench tg128 at depth 0📏 llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5; mean ± one standard deviation of 5 repetitions with warmup; pp512 and tg128 are separate short-context engine tests, not generation after a 512-token prompt; no speculative decoding or vision projector; Vulkan with KHR cooperative matrices |
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.
Blog post of 28 Oct 2026 (Valérian de Thézan de Gaussan). Means of 5 llama-bench repetitions. Not listed: Mistral Small 4 119B (32.72 tok/s at UD-Q4_K_M / 29.83 at UD-Q5_K_M) — not a current RigBase model; GLM-5.3 Flash UD-IQ1_S at 19.02 tok/s — below Q4 (~1-bit). An earlier Framework Desktop Strix Halo build on the site (geerlingguy/ai-benchmarks#21) used older models and a different source.
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/framework-desktop-strix-halo-qwen38-gemma4-valerian/)
HTML
<a href="https://getrigbase.com/builds/framework-desktop-strix-halo-qwen38-gemma4-valerian/" rel="noopener"><img src="https://getrigbase.com/badge/framework-desktop-strix-halo-qwen38-gemma4-valerian.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Strix Halo (Radeon 8060S) run Gemma 4 31B? · Can Strix Halo (Radeon 8060S) run Qwen3.8 27B?