NVIDIA Maxwell Speculation Thread

[3dilettante]

I'm hoping a reviewer with a proper understanding of the technology can do a proper OS-level inspection of where NV has made the changes. Are they sending more work to more CPU cores? Are the big improvements assuming big CPU's to help? OR is it truly an efficiency gain, and if so, why did it take this long to find that improvement?

I mean, this sort of driver revelation two years ago would've been slaughter to AMD's hopes with the 7000 series.

Even if a reviewer had an understanding of the tech, much of what the driver is going is likely hidden without reverse-engineering some of it.

Some of the biggest gains might come from bug fixing, the total war numbers seem so large that there might have been some set of large issues they fixed.

Bespoke driver optimizations can also come into play. Some preliminary reviews seem to show Mantle-enabled and benchmarking games seem to have a higher probability of getting boosts, for some reason.
SLI gets gains, which can happen because there's so much more peak to play with and because multi-GPU is by default a mess.

More intensive profiling by Nvidia could help, and a more exotic improvement would be driver detection of very similar draw call activity. The Star Swarm numbers are an example of an application that hasn't done much optimization and a general desire to do things that sacrifice efficiency with an aversion to batching and coalescing calls for a more mutable engine.
There are DX12 slides pointing to very high reuse of command sequences between frames, and likely many sequences within frames.
Something like that might be much more likely to be more generally useful, if people tried benchmarking more than a handful of games.

 

[DavidGraham]

Not bad, seems other games are affected as well, like Ghosts, sleeping dogs and Thief. StarSwarm gains are huge though.

http://www.computerbase.de/2014-04/...er/3/#diagramm-battlefield-4-nvidia-vs-mantle

 

[pjbliverpool]

Nvidia have really delivered the goods here. They've also implemented a toggle mode on the new "fake" 3D vision like the community was asking for.

 

[Alexko]

It seems that you should really use Mantle when developing your games if you want the best performance on NVIDIA hardware…

 

[DavidGraham]

Just tried the StarSwar demo myself, Win7, 660Ti OC, i7 3770, Extreme, Follow scenario:

Before :32 fps
After :59 fps !!

 

[Freemantle]

Just tried the StarSwar demo myself, Win7, 660Ti OC, i7 3770, Extreme, Follow scenario:

Before :32 fps
After :59 fps !!

Nice, no brand new low-level API required.

 

[swaaye]

Anyone try sneakiness like changing the executable's name and seeing if performance returns to previous levels???

 

[DSC]

https://forums.geforce.com/default/...feedback-thread-4-7-14-/post/4182723/#4182723

Future driver updates will bring more performance improvements, this is just only the beginning.

 

[Ailuros]

http://www.fudzilla.com/home/item/34451-nvidia-20nm-maxwell-comes-in-late-2014

If Maxwell is ever to reach between 8 and 16 Gigaflops performance as outlined in Nvidia’s roadmap, the upcoming parts need to pack more transistors. Currently the fastest single chip Geforce GTX Titan Black has 5,121 Gigaflops and we expect that an upcoming high-end GK chip will start with at least 8 Gigaflops single precision, as this is what Nvidia GPU roadmap has implied.
It will take a dual card and probably a lot of time until Nvidia’s first 20nm GM architecture reaches 12 Gigaflops.

Nothing really important to see, I just wonder if it's wise to make the same dumb mistake Is any of the fudzilla crew actually reading Fudo's stuff?

 

[silent_guy]

http://www.fudzilla.com/home/item/34451-nvidia-20nm-maxwell-comes-in-late-2014 Nothing really important to see, I just wonder if it's wise to make the same dumb mistake Is any of the fudzilla crew actually reading Fudo's stuff?

This article reminds me of high school, when we had this competition among class mates to write as many pages as possible on our history final without actually saying anything. The poor teacher had to go through mountains of BS. But at least our English wasn't as bad as this.

 

[Ailuros]

This article reminds me of high school, when we had this competition among class mates to write as many pages as possible on our history final without actually saying anything. The poor teacher had to go through mountains of BS. But at least our English wasn't as bad as this.

Nah the darned roadmap mentions that Maxwell for example will have ~12 FLOPs DP/Watt; now re-read what Fudo is rambling about.....

If Maxwell is ever to reach between 8 and 16 Gigaflops performance as outlined in Nvidia’s roadmap, the upcoming parts need to pack more transistors.

That's between 8 and 16 FLOPs DP/Watt; there's a bloody legend on the left hand side of the damn slide *sigh*

 

[homerdog]

You mean fudo is wrong and full of shit? Say it isn't so!

 

[Ailuros]

You mean fudo is wrong and full of shit? Say it isn't so!

I have a far less flattering description in mind *shrugs*

 

[iMacmatician]

At least there is an upcoming GK-based card that has 8 TFLOPS SP….

 

[pjbliverpool]

Comparisons to AMD cards in terms of TFLOPS can be misleading as NV cards are usually measured at their base clock where in reality they generally exceed their rated boost clock.

So Titan Black may be 5.12 TF base but it's 5.64 TF boost and probably running closer to 5.8 TF in reality.

I expect the 8 TF for Titan Z was also based on the base clock speed.

 

[FUDie]

Comparisons to AMD cards in terms of TFLOPS can be misleading as NV cards are usually measured at their base clock where in reality they generally exceed their rated boost clock.

So Titan Black may be 5.12 TF base but it's 5.64 TF boost and probably running closer to 5.8 TF in reality.

I expect the 8 TF for Titan Z was also based on the base clock speed.

I don't know about Titan Black, but Titan can't maintain peak boost clocks for more than a minute or so with a heavy compute workload. If you enable fast doubles, the drop off is even more significant.

 

[DSC]

http://www.geeks3d.com/20140409/asus-geforce-gtx-750-gtx750-phoc-1gd5-review/

As you can see on the benchmarks scores, the GTX 750 offers more or less the same performance than the GTX 480. But when it comes to the power consumption and temperature, the GTX 750 is just insane: the power consumption of the GTX 480 is five times the power consumption of the GTX 750. And the idle GPU temperature of the GTX 480 is the burn-in GPU temperature of the GTX 750!

This GTX 750 is a really nice product. I really appreciate the performance/watt ratio of this card:

GTX 750 = 5X POWER EFFICENCY of the GTX 480

 

[LiXiangyang]

http://www.geeks3d.com/20140409/asus-geforce-gtx-750-gtx750-phoc-1gd5-review/

In compute-bounded tasks, sure they do, however there are more compute tasks which are bounded by memory bandwidths/IO etc than bounded by computing resources.

 

[CarstenS]

It seems that you should really use Mantle when developing your games if you want the best performance on NVIDIA hardware…

This REALLY made my day! Thanks for the laugh

Comparisons to AMD cards in terms of TFLOPS can be misleading as NV cards are usually measured at their base clock where in reality they generally exceed their rated boost clock.

So Titan Black may be 5.12 TF base but it's 5.64 TF boost and probably running closer to 5.8 TF in reality.

I expect the 8 TF for Titan Z was also based on the base clock speed.

The difference between the duallies are larger than for the single-gpu model either way, so the result (usually measured in bar length) should be extrapolatable (is there such a word?)

 

[UniversalTruth]

So, some sources (PCTuning, Tyden.cz, Expreview) claim that GTX 880 will be based on 20nm GM204 with following further specs: