BM-001TESTED 2026-07-25
Qwen3-Coder-30B-A3B on Arc B60 Pro vs. dual RX 9700
Vulkan on Intel Battlemage lands within 12% of dual-RX 9700 for code generation at Q4_K_M, at roughly half the system cost. Arc is the price-performance leader for sub-$2K builds; multi-GPU AMD wins raw throughput once you accept the complexity.
vulkanq4-k-mllama.cpp b3500codingintel-arcamdvulkan30b-class
Configuration
- Model
- Qwen3-Coder-30B-A3B-Instruct
- Artifact
- Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf
- Checksum
- sha256:pending-real-checksum
- Quantization
- q4-k-m
- Context length
- 8,192 tokens
- Backend
- vulkan
- Runtime
- llama.cpp b3500
- Settings
- -ngl 99 -c 8192 -t 8 -b 512 --no-mmap
- Workload pack
- workload-pack-v1
- Author
- Edgar
- Reviewer
- Edgar
Results
100-prompt HumanEval-style coding suite plus 20-prompt long-context refactor task. 3 warmup runs, 5 measured runs averaged, variance reported.
43.8tok/s
GEN · Ray
539tok/s
PROMPT
580ms
TTFT
17.9GB
VRAM USED
| MACHINE | PROMPT tok/s | GEN tok/s | TTFT ms | VRAM GB | RAM GB | POWER W |
|---|---|---|---|---|---|---|
Jitori PC Single Arc B60 Pro, Vulkan (ANV 25.2.8). GGML_VK_VISIBLE_DEVICES=1. | 412.3 | 38.6 | 740 | 18.1 | 3.2 | 245 |
Ray Dual RX 9700, Vulkan. Single-GPU results; second card idles for this workload size. | 538.9 | 43.8 | 580 | 17.9 | 3.1 | 410 |
Quality scores
CODING
0.81/ 0-1
HumanEval pass@1
Limitations
- ◆Q4_K_M trades measurable quality for memory; pair with the Q6_K comparison (bm-003) before final decisions.
- ◆Single-GPU Ray result underuses the second RX 9700; combined-GPU scaling is a separate benchmark.
- ◆Vulkan-only for Arc per Battlemage SYCL instability; CUDA path on Victor is reported separately (bm-002).