This is amazing work - RDMA on these smaller unified memory boxes (somewhat) bridges the gap for consumers from the ~24GB 3090/4090/7900 card that are around to 128GB/256GB! Still not cheap, especially now, but... obtainable?

I do hope that apple opens up RDMA for their TB4 machines... ds4 using TB5 macs works great - but there are a lot of capable tb4 (M2/1) machines out there and afaik there's no hardware limitation preventing RDMA from working (at lower bandwidth, but with the latency gains!) on the older stuff.