AMD ULPS Is Killing Multi GPU Performance & Smoothness

[LtMatt]

I would like to see AMD give an option in CCC to disable ULPS. I also believe AMD should reccomended all review sites to disable it before benchmarking their multi gpu solutions. Single gpu's are unaffected.

Thanks to Tommy for compiling some results for me.

ULPS ON

Spoiler

There is no drawback to disabling ULPS, unless you consider an extra 10W at idle a drawback. Plenty of drawbacks to leaving it enabled though.




ULPS OFF

Spoiler

Note the extra performance with ULPS off on the minimum and average fps and the low gpu usage of gpu 2 with ULPS on. Seriously harming performance and decreasing smoothness and frame time as one gpu is not pulling its weight evenly in the usage stakes.

 

[LtMatt]

Nice find but I wouldn't go as far as saying ulps is "killing" performance.

I would consider a 17% reduction in average fps to be killing performance.

 

[LtMatt]

Does this have any effect on single gpu performance?

Single gpu's are unaffected.

ULPS is Zerocore. It shuts off the second gpu when not in use so it won't affect single gpu. Its a nice feature but it seriously harms performance and there should be an option in CCC to disable it without having to use a third party overclocking tool. Extra performance and smoothness is noticeable as soon as you disable it.

 

[LtMatt]

That's not really the point though, even if you aren't overclocking disabling it gives you extra performance, there should be an option or something put in the driver where UPLS disables itself when a full screen application is launched, not sure if it requires a restart though to disable it, I've never needed to use it.

Each time it is enabled or disabled you need to restart before the setting is applied.

Just gave this a try on my setup as I've always run with ULPS active. Using the Sleeping Dogs benchie I got:

With ULPS active -
Average: 70.8FPS
Max: 92.1FPS
Min: 43.1FPS

With ULPS de-activated:
Average: 71.2FPS
Max: 93.9FPS
Min: 44.7FPS

All at 5760x1080x60Hz, AA-High, Hi-res-On, Shadow-High, SSAO-High, V-Sync-Off, Motion Blur-On, FPS Limiter-Off

It's a bit of a difference in favour of ULPS being off, but it's not as dramatic a difference as others are seeing. Wonder if it's because of the resolution, or maybe because I've only clocked the cards to 1000/1500?

EDIT: did notice GPU2 was at 99% for both runs, with GPU1 at 95%+ the whole time too.

Sounds like it wasn't disabled.

 

[LtMatt]

I took the easy method to check if it was disabled or not - peek under the desk and see if my GPU2 fans are spinning. Fans were stopped at desktop with ULPS enabled, and were spinning at desktop with ULPS disabled.

Furry muff. Was the gpu usage constant at 99%? Because i see similar usage like in the pics in the OP with ULPS on. Gpu 1 pegged at 100% and Gpu 2 anywhere from 70-95%. I'm on single screen, so i don't know whether that has any effect or not.

 

[LtMatt]

Amen to that. I love the idea of Zerocore, its superb. But the negative effect it has on performance and to a lesser degree smoothness means its a must for being disabled.

 

[LtMatt]

Matt, I get similar results to you on Sleeping Dogs at same settings on my 2 7970's, but I am not getting 100% (more like 60-70%) usage on both gpu's.

Now I know I am going to be somewhat bottlenecked due to running a 920@4ghz, but does that seem right to you? Just having a wee play around with my cf set-up and wondering why the disparity between our respective gpu usage?

Tommy ran those benchmarks in the OP. He has a 3770k at around 4.5ghz and 2x7950's. Not sure if they were stock or overclocked though. Do you have Hyper threading enabled on your cpu? I would have thought you might be a slightly bottlenecked, but nothing drastic. If anything i guess maybe you might be a tad bottlenecked of Mhz speed. You can always try going further and see if usage improves. Generally though i find full usage on one gpu and lower usage on the other.

I will run a benchmark myself at 1440p and report back. Note at this res im much less likely to suffer any forum of cpu bottleneck. Saying that though i do have a 2700k so i do have faster IPC than you.

I'm currently running [email protected], ddr3 @2233mhz, BLK 104.7, and 7950's @1093/1563. Will report back shortly with some figures and see what my gpu usage is like with ULPS on.

 

[LtMatt]

I get terrible results with ULPS on, that could have something to do with my overclocking via BLK though.

 

[LtMatt]

Been running higher power limits since I found out how to apply it due to the boost bios being a pile of ****, if you don't want to flash your bios, you can hack the power limit and it will keep your clocks from spiking, which can give quite a hit to fps, details here:

http://forums.overclockers.co.uk/showthread.php?p=24212468#post24212468

It's very simple, I was just thorough with the instructions.

It's maybe just a lazy card though as the voltage is 1.25 in the first place.

Bios flashing is where it's at, there is an Ice-Q tbios thread here somewhere, maybe somebody can put up the link.

Btw, put your better gpu in the top slot for better performance in non CrossFire titles-if you play any.

Tommy if you use one of the custom bios created by Kaboom then +20% seems sufficient. However if you stick with the boost bios then 30% is needed i find.

 

[LtMatt]

I reckon if your cards are at stock clocks then the impact of ULPS performance hit is smaller. However if you overclock the performance hit of ULPS is massive. Tommy or anyone else would be good to see some stock comparisons.

7950's @1100/1500 ULPS off with frame pacing

Spoiler

7950's @1100/1500 ULPS on with frame pacing

Spoiler

Not entirely sure what clock or gpu usage i was getting from the second gpu with ULPS on as gpu 2 monitoring is unavailable with it enabled. I also overclock my DDR3 via BLK so im not sure if that has an effect on my performance with ULPS on as it does cause some problems with sleep mode ive noticed.

 

[LtMatt]

Can do Tommy. I couldn't complete the benchmark with ULPS on as it crashed, likely because the other gpu was stuck at stock voltage of the bios im using which is 1.175v. However with ULPS i did manage to bench safely as i could change the voltage of both cards.

CPU @4.7
2133mhz DDR3

My minimum is a bit higher and my average slightly lower. I wonder if my minimum is higher as im using some rather tight timings on my DDR3? I guess my average is lower though because im pci-e 2 8x8 vs your pci-e 3.

Spoiler

 

[LtMatt]

Cheers matt, good to know it's similar performance, maybe extra processes holding me back a touch or the memory timings like you say.

Never oc'd ram before, wouldn't know where to start tbh as I never thought ram clocking brought much to the table anyway, but in this case it might.

I vote you contact an anand guru and tell him to get on the case to Amd about the ulps performance hit, you could go down in history as the guy who pointed it out, but if Amd send you 2 9970's as thanks, ones for me.

It might not be the ram that was just a guess. Is your ram at the same speed? I tweaked the secondary timings a little after looking at one of 8packs posts, not sure how much if any difference it made though.

I don't have a twitter account so i asked Abundant Cores (Humbug) to pass it onto AMD Roy & Thracks and he did.

It would be nice to see ULPS on/off available as a feature in CCC. I'm convinced it will bring extra fps and lower frame times as well. Hopefully a review site will pick up on it and do some official testing with it on and off.

Another thing id like to see added to CCC would be the option to change flip queue size. The default setting of 3 can produce some rather unsavory input lag in twitch shooters. Nvidia users have the option to tweak it so we should as well. Of course it can be done easily in RP, but still having the option to do it in CCC would be nice.