cost per transistor has gone up slightly, also because it's new tech, but primarily because they can price it however the **** they want unless Polaris 10 is ****ing cheap and clock VERY high (meaning a huge amount of changes to the arch to allow for literally +50% clocks) I expect this is how Polaris launch party will look Spoiler
Damn after reading the highlights of the 1080 makes me want it even more. This would be a huge upgrade from my r9 290.
literally double ALU throughput, at least double geoemtry performance, but less memory bandwidth. rop count i imagine is the same TMU count maybe higher on 290
But I see past all this stuff such as double the amount of VRam, double the core clock speed when GPU boost kicks in. New Vram technology. May run cooler under load compared to my r9 290.
Just watched the presentation. Man, this is tough. The 1080 at that level of performance for a single card user like me is very titillating. Considering its availability for end of this month, its very attractive for a step up from a 980TI (going with a Founder's Ed). Then can just sell it off and upgrade to TI. Honestly was not planning to go this route and was just gonna wait out till the TI drops.. but after watching that preso, hmmm. Can't wait to see some benches of the 1080!
They went for the exact same split as with Maxwell. GP104 is 2/3rds of GP100 the geforce GP100 (GP102?) will have 3840 SPs probably with a boost of ~ 1650mhz and 8gb HBM how does ~ 14tflops sound ? THAT is actually more than double the alu throughput of a stock 980ti/titan x
This certainly looks more like the new process node. And makes it all the more difficult to believe that Polaris 10 would sport 800MHz like that one rumour table seemed to suggest. Clocking as high as possible is free performance. It's good to see Nvidia doing it again. Although I won't be buying a card of this first 14/16 generation, it's nice to see where things are going.
That graph where they show Power vs Performance for Maxwell cards vs the GTX 1080, I've extrapolated that this configuration of Pascal has 'only' 65% more performance per Watt than Maxwell: Titan X = 250W GTX 1080 = 180W Relative performance of 1080 to Titan X read from graph = 4.3/3.6 = 1.19 (or 19% more performance) GTX 1080 uses 72% (0.72) of the power of Titan X: (180W/250W = 72%) Using the above determined variables (in bold font) Extrapolated performance of GTX 1080 at same wattage as TitanX = 1/0.72*1.19 = 1.65 or 65% more performance per Watt. I'd expected the performance per Watt to be at least double that of Maxwell due to the 2 node die shrink from 28nm to 16nm (they skipped the 20nm node), and also given that it is a new architecture I thought they might have had IPC improvements too. A 65% performance increase per Watt seems a little disappointing.
Holy **** @ those clocks, all I can say lol... But also, what's this "founder's edition" all about? Maybe it has some kind of voltage unlocks so that you can reach this very high boost clocks, compared to the "regular" ones?
All I'm doing is comparing the relative gaming performance (the graph that is on Image 8 in this article) and relative power consumption of the GTX 1080 to the Titan X and the calculations I've done show that 65% increase in Performance Per Watt. I think that makes the detail of your "#ALUx2xClock/Power" irrelevant unless I'm missing your point.
Probably you already read it by now, but I will quote Hilbert just in case from the main article: "Small update: the founder edition are better binned products that should offer better overclock potential...."
I think you've misunderstood how I did my calculations unless you can be more specific. I specifically used the gaming performance graph in Image 8 of this article to do the calculations, so it does relate to gaming performance.
You used the graphs The correct way of calculating performance per watt is considering alu throughput Then if you consider another 1.5x performance (compounded) in VR you get triple the power efficiency of maxwell (in vR)
Why is a theoretical measurement (ALU throughput) more relevant than gaming performance - do you buy your GPU to look at ALU throughput numbers or to play games?! I was using the graph in Image 8 to do the calculations, I think gaming performance is more relevant than some ALU throughput. EDIT: oh yeah, if you use the VR gaming performance graph then the gains are a lot larger, but that's more niche & not adopted yet: I was using the arguably more relevant 'General' Gaming Performance graph in Image 8 of the article. Using that graph performance of Pascal is only 65% more performance per Watt than Maxwell (extrapolated calculations I did from that graph).
