Builds / AMD Radeon RX 6600M 8GB (desktop adapter, no ReBAR)
RX 6600M 8 GB without ReBAR: Qwen3.5 9B IQ4_NL 17.5 to 36.7 tok/s once llama.cpp Vulkan stops using the 256 MB BAR (fix merged 6 Nov)
Last updated:
- GPU(s)
- AMD Radeon RX 6600M 8GB (desktop adapter, no ReBAR)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 8
- CPU
- AMD Ryzen 5 5000-series
Amazon · eBay (used) · Newegg - Motherboard
- not stated (no Resizable BAR)
Amazon · eBay (used) · Newegg - RAM
- 32 GB DDR4-3200
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Debian testing, kernel 7.2.12, Mesa 26.1.6 RADV; llama.cpp master 10d4846a5, llama-bench -ngl 99 -fa 1 -p 512 -n 128 -r 5; fast rows use GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1, which selects the same path as merged PR #31658
- 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 9B | Unsloth IQ4_NL (Qwen3.5-9B-MTP-GGUF) | llama.cpp master 10d4846a5 (Vulkan, BAR path off) | 36.71 | 542 @ 512-token prompt est. TTFT @8k ≈ 15.1 s (calculated from measured prefill) | context: llama-bench tg128📏 tg128 36.71 ± 0.08 with GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1 vs 17.52 ± 0.00 on plain master (2.1x); pp512 unchanged |
| Qwen3.5 9B | Unsloth IQ4_NL (Qwen3.5-9B-MTP-GGUF) | llama.cpp master 10d4846a5 (Vulkan, default) | 17.52 | 539 @ 512-token prompt est. TTFT @8k ≈ 15.2 s (calculated from measured prefill) | context: llama-bench tg128📏 Plain master before the fix: buffers land in the 256 MiB BAR heap on a no-ReBAR system and run at PCIe speed |
| Gemma 4 E4B legacy | Unsloth QAT UD-Q4_K_XL | llama.cpp master 10d4846a5 (Vulkan, BAR path off) | 58.95 | 2,140 @ 512-token prompt est. TTFT @8k ≈ 3.8 s (calculated from measured prefill) | context: llama-bench tg128📏 tg128 58.95 ± 0.05 with the env var vs 60.43 ± 0.23 on plain master (-2.4%, host-side cost, reproduced 3 times) |
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.
User measurements posted on the llama.cpp pull request 'vulkan: detect ReBAR support, avoid BAR allocations if unavailable' (comment 5 Nov 2026; PR merged 6 Nov). On GPUs without Resizable BAR, older llama.cpp Vulkan builds could put small weight buffers in the 256 MiB BAR window, halving generation speed on some models. Builds after the merge pick the fast path automatically. A second commenter saw the same 2x (15.9 to 33.5 tok/s) on a no-ReBAR Vega 10 card.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.