I would also like to see the AI inference performance numbers of the M3 Ultra with 256 gigs of UMA memory, and I would like to see Apple do whatever they can to support developers software wise who want to network multiple M1, M2, M3, Ultra's together (with that big UMA memory foot print). It will be interesting to see what the 'M3 Ultra' 80 core results will be. From the graph on their website, it looks as if the M3 Max is in the top 12% or so of ALL benchmarks, regardless of the hardware architecture. So the M3 Max (40 cores) is an improvement over the M2 Ultra with 76 cores (3315.19). My older AMD Radeon RX 6900 XT GPU is getting 2171.32. The top NVIDIA score is 11412.38 with the GeForce RTX 4090.
The top Metal benchmark score today on their website is 3627.19 with the M3 Max with a 40 core GPU.