i have a question about raw teraflop performance. 20TF ampere doesnt convert exactly as turing. What could be the reason ? Did they gimped the architecture as teraflop count and performance gain wasnt that massive? i know teraflops doesnt matter but still i feel like a 20TF ampere card should beat down 2080ti. What went wrong in the architecture jump? or games not taking advantage of new architecture?
Ampere peak TFs include INTs into themselves since Ampere has a unified FP+INT SIMD.
Turing peak TFs doesn't include INTs into themselves since Turing has a dedicated INT SIMD.
A modern gaming rendering workload may consist of INT instructions up to 50% of all code, depending on a game or even a scene.
Thus a game which has 25% of its shader instructions as INTs will get a +25% boost on Turing not reflected in its TFs - but won't get the same boost on Ampere because these 25% INTs will run on the same h/w as flops in its case.
So the comparison of flops between them should account for that. Basically Ampere can't be 2X Turing in actual performance according to flops because its execution model of gaming code is different to Turing. It will be anywhere from 1X to 2X (per SM) in practice, depending on proportion of INT instructions in the rendering code.
That's one part of the reason. Another one is the current target h/w which games are being written for at the moment. Games target a specific balance of compute to frontend and backend throughputs in GPUs. So if a game is targeting a 1:1:1 ratio here and you have a new GPU which is 1:2:1 roughly speaking the code won't be able to take advantage of this doubling or the middle stage (additional flops) because it will still target the old ratio - and will be limited by these 1s. Generally GPUs are adding flops all the time without increasing the FE and BE anywhere nearly similar, the focus is shifting to compute for the last 20 years as shading is getting more and more complex while external bandwidths aren't changing as fast.
This is the second part of the reason - games just don't have complex enough shading which will be able to load all the flops of Ampere, yet. This will eventually change over the next couple of years and Ampere will show higher gains over Turing than it does on current gen games.
discord.gg
videocardz.com
