Apple markets a four-Mac cluster for trillion-parameter local inference

Alex Cheema's post describes an Apple marketing video that shows four M5 Ultra Macs clustered, with 2TB of unified memory at 4.8TB/s, under the line "Run trillion parameter frontier models locally." Those totals are aggregates. Apple's Mac Studio specifications list up to 512GB of unified memory and 1.2TB/s of memory bandwidth per machine, and four machines multiply out to roughly 2TB and 4.8TB/s.

That arithmetic accounts for the totals but does not measure how quickly one model runs across the machines. The bandwidth sum describes memory access inside each computer, not the speed of the connections between them, so it is not a single shared pool. A trillion-parameter model's actual performance needs a specified model configuration and a cluster benchmark, which the post does not report.