llama.cpp Underperforms on Intel iGPU in Real-World Test
nirurin · reddit · 2026-07-13
A user tested llama.cpp on an Intel iGPU (Arrow Lake), concluding that the iGPU setup is far from ideal.
Benchmark Results
- Vulkan: Basically unusable or extremely slow, peaking at around 1 tok/s.
- SYCL: The best performer; Qwen3.6 35B reached around 12 tok/s, with prefill hitting 20 tok/s, though it was unstable and sometimes froze.
- CPU: The most stable option, delivering around 14 tok/s for inference and 30–40 tok/s for prefill.
Deployment Environment
He tested various combinations on Unraid:
- CPU / Vulkan / SYCL Docker containers for llama.cpp
- LM-Studio running inside a Webtop container
- CPU / Vulkan containers for llama-swap
Conclusion
He suspects that either Intel iGPUs are genuinely not very useful for these inference tasks, or his configuration is flawed. He is asking the community if anyone has managed to run it smoothly and achieve actual performance gains.
Related event: Intel 285HX Tests Show CPU Beating iGPU for Local LLMs(2 posts)→
More from Infra
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27
- Revenue per GPU can rise as token prices fall, argues Gavin Baker — GavinSBaker · 2026-10-27
- Anthropic S-1: only $19B total compute spend since inception, overbuild warning — bookwormengr · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Grid queues make power architecture an AI-factory decision; Exowatt and Redwood offer fixes — ingliguori · 2026-10-27
- Two CVSS 9.9 Wasmtime flaws allow sandbox escape, one ARM64-only since 2020 — basedjensen · 2026-10-27