PCWatch: PS4 Technical Overview @ GDC 2013

[WereKitten]

ethomaz said:

WereKitten said:

Uh? Where are you getting this magical duplication?

Graphics and compute in parallel means more granularity in sharing the CUs, be that because of the increased number of queues as shown by the VGLeaks diagram of a few weeks ago, or by something akin to hyperthreading in CUs- i.e. a hardware assisted context switching.

The overall global limit is still 1.8TFlops; these improvements will make sure it can be used with an efficiency closer to 100%, but not magically duplicate it.

Read here... the CUs can do both at the same time without lost performance... all the 1.8TFLOPS can be used to Graphcis and Compute at the same time... on PC you need to choose in use one or other in each CU... PS4 GPU has the ability to do both in each CU at the same time... that's seem duplicate magic for me.

Or I'm geting all wrong.

Yes, that's what I read. But an ALU is an ALU... there's no magic that will make it crunch double the instructions per cycle only because you tag one as graphic and the other as compute.

The way I read that is that, going back to the vgleak document about the increased number of queues, you'll be able to keep the CUs occupied with graphic and computing tasks in parallel, without any big scale mode switch, resulting in a greater efficiency.

A Flop is a Flop... if it were like you said, those CUs would just have double the maximum theoretical Flop throughput and it would not make any sense to mark them as 1.8Tflops. They would just be indicated as 3.6Tflops because that's how many operations per second they would be able to crunch in the best scenario.