Hi, i was inquiering about the flop/compute balance in Tegra Pascal, compared to what's in the XBO/PS4 (GCN 1.0?) since you said Nvidia doesn't always outperform AMD per flop.
What I said should answer that too. Like the desktop parts, it will have higher performance per flop, or if you prefer lower flops for a given amount of performance, than any GCN architecture, since it is the same architecture as the desktop parts now (k1 and x1). A Tegra with Pascal isn't yet out, but based on the Maxwell and Kepler ones it should have the same performance to flops ratio as the desktop parts, just scaled down.
What I meant was more pre-GCN, or perhaps post-GCN, the ratios of ALU performance to shading performance to pixel fill to texture mapping is always shifting. GCN just favored compute, Kepler cut some of it out to make a more efficient gaming chip, Maxwell only brought some back. May shift again with Polaris vs Pascal.
So I guess, my point was just, "An AMD flop will always perform worse than an Nvidia flop" is too general for my liking, and nor is it inherently a bad thing as AMD favored compute, hence more flops in a given architecture for a certain graphics performance level.
Intel too, more FLOPs, and thus compute performance, than you'd expect from their IGPs graphics performance level. Like AMD, favoring compute.
So why might one want more flops than one would expect from a given performance level? Check this out: once you pushed the full 128 (!) compute sources GCN could handle without longer execution times, it could do 10x (!!!) the compute work without slowing down as Maxwell so far.
To Nvidias credit, a low amount of compute work ended up being faster on Maxwell than GCN. But once you add more compute work, GCN does it without increasing execution time, unlike Maxwell.
http://www.extremetech.com/extreme/...ading-amd-nvidia-and-dx12-what-we-know-so-far
https://forum.beyond3d.com/threads/dx12-performance-discussion-and-analysis-thread.57188/page-11
AMD went compute heavy current gen, Nvidia did not, there's nothing inherantly good or bad about an "nvidia flop outperforming an AMD flop". It just means for performance X AMD will have more compute flops at its disposal. It's like when people say GPU X does performance Z with less cores than GPU A. Well...Who cares? All that matters is performance, price, power draw, and for the GPU maker, die size.
To put my rambling more simply...GPU 1 has performance of 20 units, and 20 flops, which happen to measure compute performance. GPU 2 has 20 performance, for 30 flops. Is GPU 1 "better" just because it has higher performance per flop? Or is GPU 2 better because it has the same performance, for *also* more flops, which more measure what you can compute with it than its gaming performance?