In January 2012, the first 28 nm graphics card was released. Since then, architectural changes and new types of memory have helped increase performance and efficiency, but the process node for discrete GPUs has stayed at 28 nm for all these years. Earlier this year, AMD announced the Polaris architecture, which uses the Globalfoundries 14 nm and TSMC 16 nm processes, and NVIDIA announced the Tesla P100 processor for servers, which uses TSMC’s 16 nm process. Earlier this month, NVIDIA announced the GeForce GTX 1080 and GTX 1070, which are Pascal architecture GPUs on 16 nm, and these will be released starting later this month. Code: GTX 1080 GTX 1070 GTX 980 Ti GTX 980 Architecture Pascal Pascal Maxwell Maxwell Chip GP104 GP104 GM200 GM204 CUDA cores 2560 1920 2816 2048 Base clock 1607 MHz 1506 MHz 1000 MHz 1126 MHz Boost clock 1733 MHz 1683 MHz 1075 MHz 1216 MHz Boost GFLOPS 8872 GFLOPS 6463 GFLOPS 6054 GFLOPS 4981 GFLOPS Memory size 8 GB 8 GB 6 GB 4 GB Memory type GDDR5X GDDR5 GDDR5 GDDR5 Bus width 256-bit 256-bit 384-bit 256-bit Memory speed 10 Gbps 8 Gbps 7 Gbps 7 Gbps Bandwidth 320 GB/s 256 GB/s 336 GB/s 224 GB/s TDP 180 W 150 W 250 W 165 W Release date 05/27/2016 06/10/2016 06/01/2015 09/18/2014 Release price $599 MSRP $379 MSRP $649 $549 $699 Founder’s $449 Founder’s Edition Edition While the number of “cores” of the GTX 1080 is between that of the GTX 980 and GTX 980 Ti, they are clocked much higher in Pascal, which allows the GTX 1080 to have 1.5x the floating-point performance of the 980 Ti. The new GDDR5X memory doubles the data transfer per clock compared to GDDR5, resulting in the data rate increasing from the 8 Gbps maximum for GDDR5 to a theoretical 16 Gbps for GDDR5X. The GTX 1080’s memory speed is relatively low so it has “only” 43% more bandwidth than the GTX 980, and less than the GTX 980 Ti. However, Pascal features improved color compression over Maxwell, which should give the overall advantage to the 1080 in practice. The GTX 1070 is cut down over the GTX 1080 much more than the GTX 970 was over the GTX 980. It uses GDDR5 memory instead of GDDR5X, albeit of faster speed than the GDDR5 in Maxwell parts. There are rumors of a third GP104 part, one that fits in below the GTX 1070. As for pricing, you’ll notice that there are two prices listed for the GTX 1080 and GTX 1070 each: the normal price and a “Founder’s Edition.” The Founder’s Edition is actually the usual reference model, but with a new name and a new (higher) price. In 4K, the GTX 1080 outperforms both the GTX TITAN X and the R9 Fury X by over 30% on average, which is reasonable when going from an xx0 chip to a next generation xx4 chip. Next year I expect a consumer part based on the GP100 or the leaked GP102 codename. The full GP100 chip has 3840 SPs and a 4096-bit HBM2 interface (720 GB/s in the Tesla P100, I expect 1 TB/s or close to that in a hypothetical GeForce variant). My guess for the GP102 is 1.5x the GP104, that is, 3840 SPs and a 384-bit GDDR5X interface for ≥ 480 GB/s. With those specs, a GP102 should have similar gaming performance to a hypothetical GP100, so I wouldn’t be surprised if GP100 never made it to the desktop. AMD’s Polaris GPUs are aimed at the mainstream. You'll have to wait until Vega at the end of this year or the start of next year for AMD's next high end GPU. Sources: AnandTech: 1, 2, 3; TechPowerUp; PC Perspective
Pascal is going to be the shit tho, look at their power consumption to processing gains... those cards know how to lift. its going to be the bomb trying to overclock those babies too... 16nm might be an issue though if overclocking gets to a point that power inductive coupling becomes an issue for individual lanes lol. good times ahead, i think nvidea is going to do great and perhaps if amd can make true what they claim each company will revolutionize the industry in different directions. awesome stuff. and thank god they killed +2 gpu sli, too many quad card rigs basically being twice as expencive in power consumption and purchase price but preform about the same as a card 50% more expencive then 1 cards used in the array of 4... SLI and similar technologies was never good and should die a quick and painless death or at least be reduced to dual card setups ONLY for consumers. #SliSucks
Hoping to put one of these in the new PC I build in the next couple of weeks. Haven't had a new PC for nearly 5 years so I'm pretty excited. Going to get a 120hz+ monitor to go with it as well.
if you plan to buy in the next month, dont buy the new generation, either wait 2 months or buy prev gen.
Generally I expected maybe a little bit more from Pascal, but you can't argue with a 1070 beating the 980ti. That's probably the card I will get, once prices settle down. @Trickster what CPU will you get? Maybe some Broadwell-E?
Nah, get a 12 core processor, only 2 grand. Even better is you can grab 2 and shove them on a dual cpu motherboard. 24 CORES WITH HYPER THREADING, though shame about the 3 ghz speed. Come on, no kill like overkill.
I remember stumbling on one of his vids where he shoved some 5000$ processor that had 32 cores or something silly into the smallest computer case they could find. I don't really watch those things though, beyond some novel ideas of how to make a decent computer for cheap this kinda stuff didn't interest me. Whoops, it was a 18 core e5 2699, costs around 4000$
That's the CPU plebs like me get. Lawrence of Arabia should invest his oil money and get something higher in the foodchain.
to make server chips cheaper or for heat reasons (EXTREME overclocking, the kind done with dry ice), since a server I7 makes sense, for the rest just get a i5k and overclock it to get something the same speed as an i7k overclocked, with only downside more heat and marginally worse stats on heavy multi threaded programs.
Custom 1080s have landed! Starting from 669€ atm: https://geizhals.eu/?cat=gra16_512&xf=1439_GTX+1080#xf_top
This graph bar just seems strange, is this just for gpus? A lot of these numbers seem high, and why is a 970 using more power then a 980?
Its obviously total system power draw. The 970 is probably an overclocked version that sips more energy.