The Netyeti's Journal

Arc-GPU-Optimization-Guide

1 min read

Arc GPU Optimization Guide: Intel SYCL + Local LLM

This guide tracks the optimized local LLM setup for the Intel Core Ultra (Arc Graphics) using SYCL offloading. Reference the Discovery Log for context.

Components

  • Back-end: llama.cpp built with GGML_SYCL=ON and icpx.
  • Drivers: libze-intel-gpu (Level Zero) + Intel oneAPI 2025.3.
  • Service: llama-serve.service (User-level systemd).

Daily Usage Commands

Switching Workflows

Use the custom toggle-ai.sh script to manage memory and services:

# Start OpenClaw for agentic work
toggle-ai.sh openclaw
 
# Switch to terminal-only AI (Butterfish / Pi)
toggle-ai.sh shell
 
# See what's currently active
toggle-ai.sh status
 
# Shut it all down
toggle-ai.sh stop

High-Performance Aider

Launch Aider with the Arc-optimized backend (bypasses default Ollama):

aider-local coder   # Launches Qwen 2.5 Coder 7B
aider-local llama   # Launches Llama 3.1 8B

Troubleshooting

  • Memory Stats: If the logs show warnings about ext_intel_free_memory, ensure export ZES_ENABLE_SYSMAN=1 is set.
  • Service Logs: journalctl --user -u llama-serve.service -f
  • Device Detection: Run sycl-ls to verify the Level Zero GPU is recognized.

Maintenance

If llama.cpp needs to be updated:

  1. cd ~/llama.cpp/build
  2. source /opt/intel/oneapi/setvars.sh
  3. cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
  4. make -j$(nproc) llama-server llama-cli

Graph View

Backlinks