The video showcases an experiment testing the fastest external drive setup for Apple Silicon Macs, specifically using a Mac Mini M4 Pro with Thunderbolt 5 ports. The creator uses three Thunderbolt 5 enclosures, each housing a Samsung 9100 Pro 1TB NVMe SSD capable of PCIe Gen 5 speeds, theoretically reaching up to 14,800 MB/s read and 13,400 MB/s write speeds. The goal is to push the limits of external storage speed by combining these drives and testing their performance with file transfers and large language model (LLM) benchmarks using LM Studio, a tool for running LLMs locally.
The internal SSD of the Mac Mini already offers impressive speeds of around 6,000 MB/s, but the external setup aims to surpass this by using RAID 0 (striped) configuration across the three drives. RAID 0 splits data across all drives, maximizing throughput but sacrificing data safety since losing one drive means losing all data. The creator emphasizes that this setup is intended for temporary storage of large AI models, where speed is prioritized over data security. After configuring the RAID array, file transfer tests show astonishing speeds, with a 40 GB model file copying in about five seconds and sequential read speeds reaching nearly 20,000 MB/s, far exceeding individual Thunderbolt 5 limits.
The video also demonstrates loading large LLMs from the external RAID array into LM Studio. Models like the 32 billion parameter Quen Coder and the 120 billion parameter GPTOSS are loaded quickly, showcasing the practical benefits of the ultra-fast external storage for AI workloads. The creator runs benchmarks using LM Studio’s Python SDK to measure model load times and token generation speeds, finding that loading models from the RAID array is consistently faster than from the internal SSD. Despite the drives getting quite warm during heavy use, the performance gains are clear.
The creator discusses the trade-offs of this setup, noting that while the speed improvements are significant, the complexity, heat management, and risk of data loss make it a niche solution. This approach is particularly useful for workflows involving frequent loading and unloading of large models, such as multi-agent AI systems or pipelines processing large datasets. For typical users who only load one model at a time, the internal SSD is likely sufficient. The video serves as an exploratory experiment rather than a recommendation for everyday use.
In conclusion, the video provides a detailed look at pushing external storage speeds on Apple Silicon Macs using Thunderbolt 5 and PCIe Gen 5 SSDs in RAID 0. It highlights the potential for dramatically faster file transfers and model loading times, which can benefit AI researchers and developers working with large language models locally. The creator invites viewers to share their thoughts on such experiments, emphasizing that while this setup is unconventional and somewhat risky, it offers exciting possibilities for high-performance AI workflows.
