1.5 TB of VRAM on Mac Studio – RDMA over Thunderbolt 5 (www.jeffgeerling.com)

🤖 AI Summary
Apple has unveiled a significant advancement in high-performance computing with the introduction of RDMA (Remote Direct Memory Access) support over Thunderbolt 5 in macOS 26.2. This feature allows multiple Mac Studio machines to function as a unified memory pool, facilitating faster processing of large AI models—a crucial development as AI research demands ever-increasing computational resources. The testing conducted on a cluster of Mac Studios, boasting a combined 1.5 TB of unified memory, demonstrated impressive capabilities, achieving up to 3.7 Teraflops when performing complex computations, and enhancing performance for massive AI models like Kimi K2 Thinking. This innovation marks a noteworthy comeback for Apple in the HPC space, a realm it had largely vacated since the Xserve era. The RDMA capability notably reduces memory access latencies from 300μs to under 50μs, which can significantly enhance the performance of AI workloads. However, the reliance on Thunderbolt for networking poses limitations in scalability and usability, as current setups prevent easy connectivity among multiple Macs, and the configuration can be cumbersome. Despite these challenges, the combination of efficiency, performance, and quiet operation could position Apple's M3 Ultra Mac Studio as a formidable option for AI practitioners and researchers seeking local computing power, even as further questions about the broader applicability and future enhancements remain open.
Loading comments...
loading comments...