Important Tweak for GeForce 980 Ti SLI users

[Berryracer]

So I got my new ORIGIN PC Millennium Lately and while everything was working great. I had one major issue that made me perform around 10 formats/reinstallation of the OS trying different setups, overclock settings, and drivers.

the 353.06 and 353.30 drivers have a serious bug that causes the 980 Ti cards when in SLI mode to drop my score in in the 3DMark Ice Storm Benchmark drastically.

For example, if I perform the benchmark on a single 980 Ti, I would get around 150K to 153K which is about right for a single 980 Ti, but if I enabled SLI mode to enable my 2nd 980 Ti, rather than the score going up which is the case in other benchmarks including other benchmarks of 3DMark itself such as Fire Storm, the score actually drop by 50% giving me around 70K to 73K. I initially thought there might be a problem with my brand new system or 2nd GeForce 980 Ti (thank you nVIDIA for wasting endless hours of hard work troubleshooting my system due to your crappy drivers!!).

So today I did my almost 20th format and installed Windows 8.1 Pro, this time I tried something differently, by default, in the nVIDIA Control Panel Settings, the PhysX Processor is set to Auto Select (Recommended) [yeah right)], I set it to GeForce 980 Ti (2) for the 2nd card to be dedicated solely for PhysX and ran the benchmark, to my surprise, the score went up to 192686 !!!

So until nVIDIA fix their crappy drivers, this is the recommended method to use a Dual or more SLI setup with these 980 Ti cards.

It's not just me, here is another example I pulled off from Google:

Bought 980ti and scores are crazy low


GeForce GTX 980 Ti (353.06) [W8]

GeForce GTX 980 SLI (353.06) [W8]

GeForce GTX 980 SLI (353.06 Dedicated PhysX) [W8]

PS: With the 780M GTX in SLI mode on my Alienware 18 this problem isn't there so it seems to only be affecting the 980 Ti

 

[Ketchup]

So the question is, are the other scores inline with your SLI numbers, or are they dropping since the other card is only running PhysX?

As I recommended in the other thread, you should try running actual games with sigle/dual card and with/without the PhysX dedication to make sure you are using settings that work well for things outside a benchmark.

 

[Berryracer]

So the question is, are the other scores inline with your SLI numbers, or are they dropping since the other card is only running PhysX?

As I recommended in the other thread, you should try running actual games with sigle/dual card and with/without the PhysX dedication to make sure you are using settings that work well for things outside a benchmark.

You are right, while this method fixes the Ice Storm Benchmark, the other benchmarks went down a bit but nothing major
I did play some games and they were perfectly fine.

 

[Ketchup]

I would update that post with a score with just one card for comparison's sake.

 

[Berryracer]

I would update that post with a score with just one card for comparison's sake.

Roger that boss.

 

[Berryracer]

ooooooops, this is not a solution! After running the benchmarks with a single 980 Ti I see the results are worse with SLI!! What the heck!

 

[Berryracer]

I emailed Origin PC Support about this and here is their reply:


Greetings,




After speaking with our engineers, it very well may be due to poor driver coding. We recommend waiting until the next driver version is available and test the system further from there. If you have any other questions or concerns, please let us know and we will be more than happy to assist you.




Thank you.






ORIGIN PC Support Team

 

[Mr. Fox]

That's a good find, but one thing you may want to look at is if you dedicate one GPU in a dual GPU system to PhysX and check the box to dedicate it as a PhysX processor, SLI is DISABLED regardless of whether NVIDIA Control Panel is set to enable SLI. If you leave the box unchecked, SLI is still enabled. (See example screen shots attached to post.)

Did your score go up with the box checked or unchecked? I ask because if the box was checked, the underlying problem of the score dropping with SLI enabled is still not resolved even if the benchmark results improved because SLI was, in fact, disabled if the box was checked. This will affect all games and benchmarks. In some cases (Batman: Arkham Knight being the most recent example) a dual-GPU NVIDIA-powered system will perform best with SLI disabled and the second GPU tasked as a dedicated PhysX processor (box checked) because the game is a botched up mess that wasn't given proper attention by the software developer. You will find this to be true with some benchmarks as well.

The other thing to consider is with respect to the Ice Storm benchmark specifically, as this test isn't really designed with high performance systems in mind, much less those with multi-GPU configurations, so it may never function well in an SLI or CrossFire system. Ice Storm probably cannot tax the hardware enough to put your resources to good use and, consequentially, delivers a less impressive score due to extremely low GPU and CPU utilization. We see this in certain types of games and in really old benchmarks (think Aquamark, 3DMark05, 3DMark06, 3DMark 2001, etc.) that are not very taxing. Multi-GPU systems tend to shine the brightest under the most demanding load conditions. Considering the hardware target for Ice Storm, the emphasis is on testing the severely limited performance capabilities of those pathetic tablets, smartphones and ultra-portable devices which do not even support PhysX technology. It takes all the resources that these wimpy devices can muster to handle even very mundane DX9 graphics rendering.

http://s3.amazonaws.com/download-aws.futuremark.com/3DMark_Technical_Guide.pdf

 

[Berryracer]

That's a good find, but one thing you may want to look at is if you dedicate one GPU in a dual GPU system to PhysX and check the box to dedicate it as a PhysX processor, SLI is DISABLED regardless of whether NVIDIA Control Panel is set to enable SLI. If you leave the box unchecked, SLI is still enabled. (See example screen shots attached to post.)

Did your score go up with the box checked or unchecked? I ask because if the box was checked, the underlying problem of the score dropping with SLI enabled is still not resolved even if the benchmark results improved because SLI was, in fact, disabled if the box was checked. This will affect all games and benchmarks. In some cases (Batman: Arkham Knight being the most recent example) a dual-GPU NVIDIA-powered system will perform best with SLI disabled and the second GPU tasked as a dedicated PhysX processor (box checked) because the game is a botched up mess that wasn't given proper attention by the software developer. You will find this to be true with some benchmarks as well.

The other thing to consider is with respect to the Ice Storm benchmark specifically, as this test isn't really designed with high performance systems in mind, much less those with multi-GPU configurations, so it may never function well in an SLI or CrossFire system. Ice Storm probably cannot tax the hardware enough to put your resources to good use and, consequentially, delivers a less impressive score due to extremely low GPU and CPU utilization. We see this in certain types of games and in really old benchmarks (think Aquamark, 3DMark05, 3DMark06, 3DMark 2001, etc.) that are not very taxing. Multi-GPU systems tend to shine the brightest under the most demanding load conditions. Considering the hardware target for Ice Storm, the emphasis is on testing the severely limited performance capabilities of those pathetic tablets, smartphones and ultra-portable devices which do not even support PhysX technology. It takes all the resources that these wimpy devices can muster to handle even very mundane DX9 graphics rendering.

http://s3.amazonaws.com/download-aws.futuremark.com/3DMark_Technical_Guide.pdf

Wow, thanks a lot for this wealth of info Mr. Fox!! I never realized SLI was disabled if I set the 2nd GPU as a dedicated PhysX Processor, thanks for this.

Secondly, your second part of your post is exactly what the EVGA Support engineer told me last night, that Ice Storm is not a benchmark I shouldc consider on such a hidh end system.

But what still puzzles me is, why if I run it with a single GPU the score goes up from 73K to 153K and above ?

Also, why do my 780M GTX in SLI on my Alienware 18 not suffer from the same drop in performance? I think I know the answer but just checking, it's because of these crappy nVIDIA drivers that we have been cursed with month after month for a long time now.

Gone are the good old stable days of nVIDIA Drivers.... It's been a month since the release of 353.30 and I check desperately more than 40 tines a day on the nVIDIA website awaiting a new driver that will unleash the full power of those 2 980 Ti's

 

[Mr. Fox]

Wow, thanks a lot for this wealth of info Mr. Fox!! I never realized SLI was disabled if I set the 2nd GPU as a dedicated PhysX Processor, thanks for this.

Secondly, your second part of your post is exactly what the EVGA Support engineer told me last night, that Ice Storm is not a benchmark I shouldc consider on such a hidh end system.

But what still puzzles me is, why if I run it with a single GPU the score goes up from 73K to 153K and above ?

Also, why do my 780M GTX in SLI on my Alienware 18 not suffer from the same drop in performance? I think I know the answer but just checking, it's because of these crappy nVIDIA drivers that we have been cursed with month after month for a long time now.

Gone are the good old stable days of nVIDIA Drivers.... It's been a month since the release of 353.30 and I check desperately more than 40 tines a day on the nVIDIA website awaiting a new driver that will unleash the full power of those 2 980 Ti's

Yeah, every driver NVIDIA has released since February has been a piece of garbage that causes SEVERE throttling on my systems with Kepler SLI. With SLI disabled everything works fine, but with SLI enabled they throttle to 405MHz core and 400MHz memory under heavy load. If I use 345.20 or older drivers everything functions flawlessly. It's truly a sin and crime how bad they have messed their drivers lately.

 

[Grubbernaught]

Have you checked hw monitor to ensure your sli is actually leaving idle clocks while running this bench?

Blaming "crappy drivers" for a lower than expected score on a bench designed for mobile phones is like blaming "crappy drivers" for not optimising for Farmville.

Use firestrike extreme or modern games at a minimum for benching 980ti sli