Nvidia Pascal Announcement

I'm not sure that the chosen clocks are in the sweet spot for efficiency.

In terms of TDP (W)/Billion transistors:
980 165/5.2 = 32
1080 180/7.2 = 25
980 Ti 250/8.0 = 31

Compute-only, using HBM, has significant IO and not fully enabled, but included for reference
GP100 300/15.3 = 20

It's true that even though the process does promise 2x more power efficiency at the same level of circuit performance, there can be confounding factors like memory and non-ASIC power consumption.
I tried to adjust for the memory subsystem. Using some GDDR5 numbers from Fury's memory comparison, and roughly in line with some earlier percentages of power budget for memory being 20-30% (split the difference), I used the Ti's wider GDDR5 bus to derive a value for the interface's cost of 63W for a 384-bit bus.
The first three with GDDR5-ish interfaces (adjusted for width) go 24,19, 23. There is an assumption that the GDDR5X bus is more in line with the 980's 256-bit GDDR5 bus in power consumption than the 980 Ti's 384-bit bus. Using the latter's value puts the 1080 at a better ~16, although that might not be flattering for GDDR5X.

 

16FF is already pretty mature and Pascal is far from being conservative in terms of perf/W and perf/mm2 and it does so at the same time. Remember the TSMC 16FF+ page we used to reference here? The one that talked about X% perf increase OR Y% power decrease. The 1080 is not quite AND, but it seems to be going a long way if Nvidia's stated numbers are holding up.

What expectations do you have about Volta?
I have none other than that it will be faster with better power efficiency. But I don't think we'll see the kind of improvements that we saw with Maxwell or Pascal. You can only go for the kill so many times in terms of arch improvement, diminishing returns and all that, and the transition to FinFET is behind us now.

 

GDDR5X defaults to 1,35 volt, GDDR5 to 1,5 volt. Maybe factoring that in? Quadratically as for ASICs?

 

With voltage squared and linear with transfer rate, that should put the 1080's bus closer to the 980 than the 980 Ti's.

 

That's what calc.exe said as well. Don't know how 5X really behaves, though.

 

According to the table, it isn't. The overclocked 1080 is 3x as fast, but stock to stock it is about 2.43x as fast as a 970.

 

True, but I intend to get an OC version And when comparing to my current 670, even a stock 970 is like lightening.

 

So one real-world game performance result we have for 1080 albeit still just summary info goes back to the announcement last week.
They showed Doom obviously using beta drivers with Vulkan and the game on absolute top settings (nightmare) running between 130fps to a brief peak near 190fps.
pcgameshardware has done a recent review of Doom and the results are interesting even at 1920x1080 (which was the presentations setting as well); they used the setting below nightmare so less shadow detail.
In their test an AIB 980ti Palit Super Jetstream had a minimum 125 and average 158.7

So that is a surprising result, OK the 1080 was using Vulkan but considering the optimisation benefits are yet to translate to healthy boost in DX12 for NVIDIA this still looks impressive (or at least Vulkan is working better than DX12 lol), especially as the 1080FE was on the very top setting and the 980ti was a notch below and also 1080FE on beta drivers.
Fingers crossed pcgameshardware will repeat their test once the Vulkan patch is rolled out for a better comparison.
http://www.pcgameshardware.de/Doom-2016-Spiel-56369/Specials/Benchmark-Test-1195242/

But looks like it should raise some eyebrows regarding initial real-world result for Founders Edition 1080, anyway better information IMO than any supposed "leaked" results popping up on the internet.
Cheers

 

Average 158.7 fps. Just sayin'.

 

Thanks,
was on autopilot when typed max

Edit:
I should also had said worth checking out the reviews/comments just how much overhead setting it all to max/nightmare visuals has.
Even in the pcgameshardware review they have a quick click tab showing 76fps for ultra and 56fps for nightmare - this is a quick and dirty comparison they provided just to show how much it can hurt.
Also need over 6GB VRAM it seems according to the review for max nightmare settings.

Cheers

 

GeForce GTX 1080 Overclocking - Extremely high cycle rates are possible up to 2.5 GHz?
http://www.tweakpc.de/news/36992/ge...ng-enorm-hohe-taktraten-bis-2-5-ghz-moeglich/

 

Hmmm, according to that link. It's running 381 MHz faster than the non-overclocked one. Or ~22% faster on the core clock. Still impressive to get 24% more performance somehow from a 22% core overclock. Quantum physics? I'm taking it with a huge grain of salt. It's even less believable if it truly was only 14% higher.

Regards,
SB

 

Geforce GTX 1080: What the Founders Edition of previous reference designs differ - Interview
http://www.pcgameshardware.de/Nvidi...ls/Founders-Edition-Geforce-GTX-1080-1195206/

 

So a portion of the better efficiency isn't due to on-die optimizations, but an optimized and horizontally upscaled voltage regulator setup with a lower impedance and hence smaller voltage swings and less overall noise in that.
(~10% relative gains in efficiency? Possibly more if this stabilized power supply has follow-up effects on-die, as the plotted efficiency appears to be only voltage regulator efficiency.)

The rest is just praising the optimized airflow of the FEs radiator design at the expense of shielding everything, and reasoning about the up-charge.

 

A huge grain of salt is indeed appropriate. Running that benchmark 1.5x faster than Titan-X, all with memory at 320GB/s vs 336GB/s.
Questionable, unless there is again a very big improvement in frame compression.

 

First waterblock compatible with 1080 from BYKSI (less than $100):


http://videocardz.com/59916/byksi-announces-first-waterblock-for-geforce-gtx-1080

and on a side note, custom watercooled 1080 boards will be announced at Computex with 2.5GHz clocks

 

By the time we are at Computex we'll be at 3Ghz

 

GTX 1080 exposes 20 Compute Units in OpenCL. So CC6.1 is like Maxwell 128SPs per SM?

 

That's weird. Any chance the API is enumerating single TPC as a multi-processor, since Pascal pairs two SMs into a TPC?

 

GT200 GTX 260 with TPC-Clusters has 27 OpenCL Compute Units (9*3*8SPs = 216SPs).