Is there some method behind the seemingly random selection of discrete GPUs charted for comparison? I don’t think I’ve seen a GTX 980 compared to anything inside the last decade, what made it crop up here?

Keen to see the follow up whenever they get their hands on the M3 GPUs, as we should be expecting some increase in L1 capacity to service the raytracing hardware.

A question for anyone more familiar with GPU architecture: what limits the latency on the scratchpad memory?

For comparison, AMD’s 7950X gets 84.06 ns using 4 KB pages over a 1 GB array. AMD’s Phoenix suffers from higher latency and hits 126.7 ns with the same 1 GB test size.

It’s a bit annoying when he wrote “for comparison” but these are of UNKOWN quality. 84.06ns on what memory configuration? Suffers from 126.7ns on what configuration?

Previously graph shows bandwidth of less than 100GB/s, and it looks like 85-90GB/s. It could be LPDDR5-5500 or DDR5-5600 of various timings. Again no mention if these are even on the same system. 126ns sounds like JEDEC DDR5-5600B. So if the desktop is using something like DDR5-6000 CL36 then that latency really has nothing to do with Phoenix.