Metal Sandbox schedules GPU coding agents across Apple Silicon Macs
Original title: Metal Sandbox: Isolated Execution and Command Scheduling for GPU Coding Agents on Apple Silicon
On four Mac mini M4 nodes, command throughput rose 1.68×–2.05× and mean tool waiting fell 18.8%.
Researchers present Metal Sandbox, infrastructure for agents that generate Apple Metal kernels. It combines restricted per-instance files, native command execution, and scheduling across Macs so parallel commands run together while serial commands share one slot per node and waiting work can migrate. Harbor integration covers task execution, verification, and rollout collection. The project is open-source on GitHub.
Unconfirmed: Reported gains assume continuously replenished sessions and unchanged command durations.
- Controlled tests used four Mac mini M4 nodes under concentrated demand.
- Sixteen static traces imply a 2.14× sustained-throughput gain at four sessions per node.
- Replayed traces cut batch elapsed time by 22.3% under dynamic scheduling.
- Eligible waiting commands can move to free Macs and return results to the original agent.
- Vendors
- Apple
- Products
- Metal SandboxHarbor
From the original article
1 min read at arxiv.orgOn-device inference requires kernels that make effective use of local accelerators. Efficient training and evaluation of models that generate Apple Metal kernels benefit from native macOS tools and concurrent agent sessions. These sessions alternate between model requests, development commands, and GPU execution, leaving GPU capacity idle for much of a rollout.
By Haozhe Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.