Builds / AMD Radeon 890M iGPU (Ryzen AI 9 HX 370, RDNA 3.5, shared me
Radeon 890M iGPU (Ryzen AI 9 HX 370): Qwen3.5 35B-A3B Q4_K_M at 24.8 tok/s on llama.cpp Vulkan b10830, 20.3 on b12387 (tg regression report)
Last updated:
- GPU(s)
- AMD Radeon 890M iGPU (Ryzen AI 9 HX 370, RDNA 3.5, shared memory)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 48
- CPU
- AMD Ryzen AI 9 HX 370 (12C/24T Zen 5)
Amazon · eBay (used) · Newegg - Motherboard
- not stated
Amazon · eBay (used) · Newegg - RAM
- 62 GiB system RAM (GTT raised to 48 GiB, 512 MiB BIOS carve-out)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Arch Linux, kernel 7.2.11, Mesa 26.2.2 RADV, GCC 16.2.1; llama-bench, identical build configs; vram_gb here is the GTT limit the GPU may use
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.5 35B-A3B legacy | Unsloth UD-Q4_K_M | llama.cpp b10830 465e49b9c (Vulkan, RADV) | 24.8 | 251 @ 512-token prompt est. TTFT @8k ≈ 32.6 s (calculated from measured prefill) | context: llama-bench tg128📏 llama-bench -ngl 99: tg128 24.80 ± 0.04 (ngl 40 22.98, ngl 20 17.66, ngl 0 14.27) |
| Qwen3.5 35B-A3B legacy | Unsloth UD-Q4_K_M | llama.cpp b12387 a956a852c (Vulkan, RADV) | 20.32 | 293 @ 512-token prompt est. TTFT @8k ≈ 28.0 s (calculated from measured prefill) | context: llama-bench tg128📏 Same machine and command on the newer build: tg128 20.32 ± 0.06 (-18.1%), pp512 up 16.6%; Arch llama-cpp 0.9.0 package (b12110, asserts on) 21.45 |
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.
Reporter's own measurements in an open llama.cpp bug report (14 Nov 2026): raw token generation on the 890M iGPU dropped ~18% between September (b10830) and November (b12387) builds while prompt processing got faster. If you run a Radeon 780M/890M-class iGPU, the older build is faster for chat-length generation until this is resolved.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.