Are there some new nvidia drivers coming

Kaapstad · 5 Jun 2013 at 21:22

As in the Title

I was looking at 3dmark11 results earlier and found this

http://www.3dmark.com/3dm11/6664924

Check out the driver on it, if I am reading it right it says 320.26 ?

andybird123 · 5 Jun 2013 at 21:44

could be some testing going on in house?

it certainly doesn't appear to be available for download anywhere

James J · 5 Jun 2013 at 21:45

Theres a new CUDA version (5.5) that comes bundled with a new driver of 320.20, yet the 320.26 are the next beta. Not sure whos jumped the gun there though....

Kaapstad · 5 Jun 2013 at 21:47

andybird123 said:
could be some testing going on in house?

it certainly doesn't appear to be available for download anywhere

Where it was a very high score, I thought it might be one of the pro benchers who could have access to something like that.

humbug · 5 Jun 2013 at 21:49

P18010 would put that 780 much faster than a GTX Titan running higher clocks than it was apparently running.

James J · 5 Jun 2013 at 22:01

humbug said:
P18010 would put that 780 much faster than a GTX Titan running higher clocks than it was apparently running.

The GK110 core has a LOT of performance left in it that drivers will unlock.

bru · 5 Jun 2013 at 22:10

Well AMD squeezed out a driver that gave a nice solid boost to performance, so there is no reason why Nvidia cannot do the same.

Kaapstad · 5 Jun 2013 at 22:16

humbug said:
P18010 would put that 780 much faster than a GTX Titan running higher clocks than it was apparently running.

3dmark11 is infamous for getting clocks wrong on highly overclocked systems.

The real question is not what clocks they were running or even the score, but where did that driver come from ?

bru · 5 Jun 2013 at 22:33

New driver for the 760ti perhaps?

Isn't it due on the 9th?

James J · 5 Jun 2013 at 23:33

Kaapstad said:
3dmark11 is infamous for getting clocks wrong on highly overclocked systems.

The real question is not what clocks they were running or even the score, but where did that driver come from ?

Theres many different none public drivers available to testers. Its more than likely one of them. Manuel G sent me some in the past due to specific issues with SLI on my old 580 3gb cards, which worked a treat.

humbug · 5 Jun 2013 at 23:43

James J said:
The GK110 core has a LOT of performance left in it that drivers will unlock.

No I don't think so, the performance difference to the GTX 680 is about right for how much GPU it is.

Some reviewers predicted this and that, none of it came true because they got their numbers wrong, I even pointed that out before the Titan got released, my numbers suggested + ~70% at best clock for clock, it turned out to be +~50 at lower clocks.

Some reviewers had 100% scaling to 100% of the extra available SP's, it does not work like that, + 100% of SP's = 60% extra performance.

Probably a mixture of not having a clue what they are talking about, or wilfully hyping it up to gain hits.

Boldie007 · 6 Jun 2013 at 08:23

Wow, I want those drivers

!

Gregster · 6 Jun 2013 at 09:12

If those drivers give those kind of scores....

Kaapstad · 6 Jun 2013 at 11:22

humbug said:
No I don't think so, the performance difference to the GTX 680 is about right for how much GPU it is.

Some reviewers predicted this and that, none of it came true because they got their numbers wrong, I even pointed that out before the Titan got released, my numbers suggested + ~70% at best clock for clock, it turned out to be +~50 at lower clocks.

Some reviewers had 100% scaling to 100% of the extra available SP's, it does not work like that, + 100% of SP's = 60% extra performance.

Probably a mixture of not having a clue what they are talking about, or wilfully hyping it up to gain hits.

I don't think there is a lot left to get out of the Titans either.

Having said that the number SMX modules available do scale very well.

If you compare a Titan to a GTX 690 for SMX modules it comes down to 14 vs 16 (2 x 8 on the GTX 690).

An air cooled Titan overclocked has the same performance as an air cooled GTX 690 @stock in Heaven 4 (they both score just over 1600)

I have done these two pics to show the clock speeds in GPUz

GTX 690s @stock in Heaven 4

Single Titan overclocked in Heaven 4

As you can see the Titan is running about 100mhz higher than the GTX 690s.

My point is 14 SMX modules running an extra 100mhz higher give the same score in Heaven 4 as 16 SMX modules at the lower speed. When it comes to SMX modules and MHz they scale pretty good.

I know in the GTX 690 pic I am running 2 of them but it was not about getting a score, it was about showing the clock speeds they run @stock.

Gregster · 6 Jun 2013 at 11:29

Good info Kaap thanks.

melmac · 6 Jun 2013 at 12:15

James J said:
The GK110 core has a LOT of performance left in it that drivers will unlock.

bru said:
Well AMD squeezed out a driver that gave a nice solid boost to performance, so there is no reason why Nvidia cannot do the same.

I think most people would disagree. The GK110 is just kepler and even then Kepler is just an evolution of Fermi. If nvidia did have a driver that made a big difference to performance they would have released it already for the 6 series cards. That's not to say that there won't be game specific improvements as usual.

GCN was completely new and that's why they were able to get a big boost performance. Just like the release of fermi (the 470/480), there was a driver that came out a couple of months later that offered a big boost to performance.

Gregster · 6 Jun 2013 at 12:19

Duff-Man (whom I respect) said he doubts there would be a big increase like the 12.11 drivers, so I am a little optimistic.

humbug · 6 Jun 2013 at 12:21

Kaapstad said:
I don't think there is a lot left to get out of the Titans either.

Having said that the number SMX modules available do scale very well.

If you compare a Titan to a GTX 690 for SMX modules it comes down to 14 vs 16 (2 x 8 on the GTX 690).

An air cooled Titan overclocked has the same performance as an air cooled GTX 690 @stock in Heaven 4 (they both score just over 1600)

I have done these two pics to show the clock speeds in GPUz

GTX 690s @stock in Heaven 4

Single Titan overclocked in Heaven 4

As you can see the Titan is running about 100mhz higher than the GTX 690s.

My point is 14 SMX modules running an extra 100mhz higher give the same score in Heaven 4 as 16 SMX modules at the lower speed. When it comes to SMX modules and MHz they scale pretty good.

I know in the GTX 690 pic I am running 2 of them but it was not about getting a score, it was about showing the clock speeds they run @stock.

I can see it in the Vally thread

#3161 - GTXTitan @ 1011 / 1666 - Kaapstad

#3980 - 690 @ 1180/1750 - Kaapstad

Titan: 2688 SP's / 384Bit / 48 ROP's #3161 = 100%

690: 3072 SP's (115%) / 512Bit (133%) / 64 ROP's (133%) / (115% Mhz) #3980 = 125%

Add in the extra memory bandwidth, ROP's and 15% higher clock on the 690. Its right

Another way you can look at it.

#2028 - 680 @ 1086 / 1800 / 1180 - Kaapstad

#3161 - GTXTitan @ 1011 / 1666 - Kaapstad

680: 1536 SP's / 256Bit / 32 ROP's #2028 = 100%

Titan: 2688 SP's (175%) / 384Bit (133%) / 48 ROP's (150%) / (105%Mhz) #3161 = 155%

Nuwidol · 6 Jun 2013 at 15:39

humbug said:
I can see it in the Vally thread

#3161 - GTXTitan @ 1011 / 1666 - Kaapstad

#3980 - 690 @ 1180/1750 - Kaapstad

Titan: 2688 SP's / 384Bit / 48 ROP's #3161 = 100%

690: 3072 SP's (115%) / 512Bit (133%) / 64 ROP's (133%) / (115% Mhz) #3980 = 125%

Add in the extra memory bandwidth, ROP's and 15% higher clock on the 690. Its right

Another way you can look at it.

#2028 - 680 @ 1086 / 1800 / 1180 - Kaapstad

#3161 - GTXTitan @ 1011 / 1666 - Kaapstad

680: 1536 SP's / 256Bit / 32 ROP's #2028 = 100%

Titan: 2688 SP's (175%) / 384Bit (133%) / 48 ROP's (150%) / (105%Mhz) #3161 = 155%

Yeah, was just about to post the same.... :confused:

Rroff · 6 Jun 2013 at 16:43

My gut instinct is we aren't seeing the full performance GK110 can bring to gaming... IMO we should be seeing closer to that ~70% (but at the current clocks rather than clock for clock) that humbug originally mentioned not the ~50% it is - not sure if I can explain this very well...

When your dealing with crunching through gigabytes of data the optimal approach is often to batch up large amounts of data at once, delay/reorder some operations to get the best long term performance which is great for plowing through lots of data but not so optimal for typical gaming scenarios where you want to quickly process smaller amounts of data - in a simplistic sense this is why AMD's old vliw architecture had such high theoretical performance and does so well at some things but struggles to bring that level of performance to gaming.

I think we are seeing something similiar with Titan, it might be that some level of performance is unavoidably lost by dispatching data sub-optimally for game type processing on a compute focused design - its possible tho they still haven't fully optimised to get the best out of it and in shader heavy games/benchmarks we are likely to see upto ~20% increases (upto 60% in context with figures humbug mentioned) with future drivers.