What Thunderbolt versions does this skill cover?

It provides technical specifications and effective bandwidth comparisons for Thunderbolt 1 through the latest Thunderbolt 5 standard.

Does this skill help with Apple Silicon eGPU support?

Yes, it includes specific documentation on Apple Silicon limitations, explaining that while hardware tunnels PCIe, custom drivers are required for compute tasks as macOS lacks official rendering support.

How are GPU disconnections handled?

It outlines a migration flow where a device failure triggers a state checkpoint, followed by selecting a target device (like an iGPU) and restoring the compute state to resume execution.

What is the difference between compute-bound and bandwidth-bound eGPU workloads?

Compute-bound tasks (like matrix multiplication) perform at 85-95% of internal speeds, while bandwidth-bound tasks (frequent CPU-GPU transfers) may see significant performance drops due to Thunderbolt overhead.

eGPU Infrastructure & Management

Name: eGPU Infrastructure & Management
Author: plurigrid

byplurigrid

•

クラウドインフラストラクチャ

Manages external GPU lifecycles, hardware detection, and workload migration strategies for high-performance compute environments.

The eGPU skill provides deep technical domain knowledge for integrating and managing external GPUs connected via Thunderbolt. It includes detailed bandwidth analysis for Thunderbolt 1-5, hardware detection patterns for macOS (IOKit) and Linux (udev/DRM), and implementation guidance for GPU hotplugging. It specifically supports the Basin infrastructure, offering strategies for workload migration and checkpointing to ensure compute continuity when a GPU is disconnected or fails. This is essential for developers building hardware-aware applications, GPU-accelerated clusters, or multi-GPU routing systems.

主な機能

01Cross-platform hardware detection for macOS IOKit and Linux udev/DRM

02Thunderbolt mesh networking for peer-to-peer GPU communication

03GPU workload migration strategies including Eager, Lazy, and Preemptive

048 GitHub stars

05Thunderbolt bandwidth math and PCIe tunneling specifications (TB1-TB5)

06Integration with Basin's EGpuManager and WorkloadMigrator Rust crates

ユースケース

01Implementing robust GPU failover and compute offloading for mobile workstations

02Developing drivers or system tools for automated eGPU hotplug detection and recovery

03Optimizing data transfer bottlenecks by analyzing Thunderbolt bus overhead

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi egpu

For use in Claude.ai and ChatGPT

Download Skill