JohnTdi/Qwen3.5-Unsloth-GGUF-R9700-Benchmark

Qwen3.5 GGUF Models — RDNA4 R9700 Benchmark

Exact GGUF files used in the Comprehensive LLM Benchmark on AMD Radeon AI PRO R9700 (RDNA4).

These files are provided so anyone with an R9700 can reproduce the results exactly.

Note: The 35B file was downloaded from unsloth/Qwen3.5-35B-A3B-GGUF on 2025-02-25. Unsloth has since updated the file — the current version on their repo is larger (20.71 GiB vs 18.34 GiB). This repo preserves the original file used in the benchmark.

Models

Model	Type	Total Params	Active/Token	File Size	Quantization	Source
Qwen3.5-35B-A3B	MoE	34.66B	~3.5B	18.34 GiB	UD-Q4_K_XL (file_type=Q4_K_M)	unsloth
Qwen3.5-27B	Dense	26.90B	26.90B	15.59 GiB	Q4_K_M	unsloth

System Configuration

Component	Details
GPU	AMD Radeon AI PRO R9700 (gfx1201, RDNA4, 32 GB GDDR6, 64 CUs)
Memory bandwidth	640 GB/s (MCLK 1258 MHz)
PCIe	PCIe 5.0 x16, 32 GT/s
CPU	AMD Ryzen 9 9900X 12-Core
RAM	64 GB DDR5
OS	Ubuntu 24.04.4 LTS, Kernel 6.19.8
Mesa (RADV)	25.2.8
llama.cpp	commit dc8d14c58 (build 8554)

Best Results

Qwen3.5-35B-A3B (MoE)

Metric	RADV	AMDVLK
Best decode	149.5 t/s	163.7 t/s (gfx+rm_kq=1)
Best prefill pp2048	3,075 t/s (ub=2048)	2,170 t/s

Qwen3.5-27B (Dense)

Metric	RADV	AMDVLK
Best decode	32.5 t/s (ASPM perf)	33.2 t/s (ASPM perf)
Best prefill pp2048	993 t/s (Mesa 25.3.6)	207 t/s

Key Findings

• rm_kq=1 is the single most impactful code change: +1% RADV, +2% AMDVLK MoE, +13% AMDVLK dense
• PCIe ASPM=performance gives +10.8% dense decode on RADV
• gfx queue helps AMDVLK MoE (+4.7%) but hurts AMDVLK dense (-8%)
• RADV wins overall (best prefill, competitive decode)
• Dense models reach 79-83% bandwidth utilization with optimizations

Full Results

See BENCHMARK_COMPREHENSIVE_R9700.md for the complete benchmark with 50+ configurations tested.

License

Apache 2.0 (following the original model license).

Credits

• Models quantized by Unsloth
• Base models by Qwen