drkohler said:
Well, they did tell the story, in a convoluted way, when readiing between the lines... Notice how they laud the 6% clock speed increase over using all 14 CUs (notice how the engineer "conveniently" mentiones the false 14+4 CU rumour of the PS4 gpu). There are two reasons why they did/had to go the way they go now. The first reason is the yield issue. The yield must have gone into the cellar when they enabled all CUs. The second reason is the gpu simply choked on insufficient ROPs and TMUs. A speed bump increases speed not only to the CUs, but also to whole rest of the gpu. And I was reading several times over the magical 204/218GB/s esram bandwidth. I still lhave no idea at all where they get their magic from. We know they have four 256bit memory controllers into esram, and a cross bar setup for the memory management so that all additional dma controllers can access the crossbar seemingly simultaneously. To me, that leads to very high bandwidths INTO THE CROSSBAR, but not into the esram. There are still four memory controllers for the esram, so there is no way you can go higher than 109GB/s into esram (with simultaneous read/writes into different blocks). the talk about read-modify-write cycles combined with a second 1024bit read bus is not enlightening either at the moment.. |
Ah, good point. There's no point increasing the CUs if the GPU is limited by the ROPs and fill-rate. I guess he wouldn't want to come out and say that though.
The esram comments confused me as well. I was actually hoping you'd post and clarify it but it sounds like you're also trying to make sense of the comments!
