Grand Theft Auto V Multi-GPU Performance Review Part 2 @ [H]

[FrgMstr]

Grand Theft Auto V Multi-GPU Performance Review Part 2 - This is Part 2 of our full evaluation of Grand Theft Auto V's video card gaming performance. In this part we dive into NVIDIA SLI and AMD CrossFire highest playable gameplay settings and apples-to-apples at 1440p and 4K resolutions. We find out just what it takes to get the most out of GTA V at its highest settings.

 

[TurboGLH]

Thanks guys. Matches up pretty closely with my results on my o/c'd 290 xfire setup.

 

[D1RTYD1Z619]

Great review. I wonder what 2 295x2s in 4k maxed out would pull? Vram usage, and power usage would be good to know as well.

 

[chenw]

Many thanks for this work!

Titan X SLI at 1440p is definitely surprising to say the least, and I am not surprised at 970 SLI struggling at 4k (it always has with AAA's anyway, no reason to expect it any different).

If CPU is indeed a bottleneck with Titan X SLI, would it be reasonable to expect that not even 5960X can completely alleviate that?

 

[bigbeard86]

if the 970 only has 3.5 gigs of vram, how is it possible to play at max settings on 1440p?

I use 290s in crossfire, and it uses 3.6 gigs vram at 1080p with all maxed, except msaa at 4x.

 

[chenw]

That is probably the most misconcepted fact about 970...

It has 4GB VRAM, segmented sure, but it still has 4GB.

Also it is based on different architecture, it might have used VRAM a little differently.

 

[phorkz]

I thought they may have used 2xmsaa for the baseline "max settings".

 

[chenw]

Aha, on 1440p page, it stated that the highest playable settings for each setup, all of them uses FXAA.

Might account for the lower VRAM usage

EDIT:

Kyle, any chance you can clarify what kind of AA, if any, were used in your performance tests?

The review stated FXAA in the beginning of each resolution, but rest of the article is completely silent on the matter.

 

[Spacy9]

I'm trying to figure out what you did different in the 4K section. I have 2 Titan Xs in SLI and running the game with everything maxed out except MSAA is off and Grass is on High. The game shows I'm using 8054 out of 24547? If I turn on MSAA and Grass to Ultra I'm up to almost 10,000 used?

Should this number be divided by 1/2 (1/2 per card)?

Thanks!

 

[phorkz]

Aha, on 1440p page, it stated that the highest playable settings for each setup, all of them uses FXAA.

Might account for the lower VRAM usage

EDIT:

Kyle, any chance you can clarify what kind of AA, if any, were used in your performance tests?

The review stated FXAA in the beginning of each resolution, but rest of the article is completely silent on the matter.

Well they say at the start, that any mention of the "maxed out" settings will use exactly what the screen shots show, so it has to be FXAA.

I didn't think alot of people would be using FXAA due to the blurriness it creates, though i may be wrong i guess, but i am not sure saying that 980 SLI is great for 2560x1440p at "maxed" settings is completely correct, change the AA to anything but FXAA and you will more than likely go over the VRAm buffer at the mentioned settings.

I am seeing 4.5gb+ VRAM usage with the same setiings using 2x MSAA(MFAA enabled) on my Titan X

I'm trying to figure out what you did different in the 4K section. I have 2 Titan Xs in SLI and running the game with everything maxed out except MSAA is off and Grass is on High. The game shows I'm using 8054 out of 24547? If I turn on MSAA and Grass to Ultra I'm up to almost 10,000 used?

Should this number be divided by 1/2 (1/2 per card)?

Thanks!

Yeah, you have to halve that, best to use a program like MSI Afterburner or something to run an on screen display while actually playing the game to see you VRAM usage anyway.

 

[Spacy9]

Yeah, you have to halve that, best to use a program like MSI Afterburner or something to run an on screen display while actually playing the game to see you VRAM usage anyway.

Thanks, this is the first time I've had an SLI setup, and I just put the second card in yesterday. I'm a noobie again

 

[Cataulin]

This is gonna sound silly but what about CF/SLI for older cards like 660's and a 6990's? everywhere I look the oldest card they used was around 750/60.

 

[x3sphere]

Well they say at the start, that any mention of the "maxed out" settings will use exactly what the screen shots show, so it has to be FXAA.

I didn't think alot of people would be using FXAA due to the blurriness it creates, though i may be wrong i guess, but i am not sure saying that 980 SLI is great for 2560x1440p at "maxed" settings is completely correct, change the AA to anything but FXAA and you will more than likely go over the VRAm buffer at the mentioned settings.

I am seeing 4.5gb+ VRAM usage with the same setiings using 2x MSAA(MFAA enabled) on my Titan X

I'm with you on that, MSAA is the real VRAM killer in this game, as soon as you turn it on VRAM usage skyrockets.

With all the intricate detail in GTA V, FXAA really hurts image quality IMO - can't stand the blur.

 

[AllBlackFan]

I know this is a "maxed out" test, but in-line with your real-world testing philosophy, I have to mention that the grass setting is having a massive impact on performance with a minor upside image quality wise imho.

Comparison

 

[Brent_Justice]

Feature performance will be shown in the next part.

 

[TMCM]

hrmmmm I'm waiting for this to go on sale but I wonder if my x-fire'd 7870's w/2GB will be able to play this @1080p on ultra or very high settings

 

[Supermr2]

I'm sure this has been asked before but is you test rig ever going to change CPU's. My 970 sli numbers seem to be higher than the articles.

 

[Armenius]

if the 970 only has 3.5 gigs of vram, how is it possible to play at max settings on 1440p?

I use 290s in crossfire, and it uses 3.6 gigs vram at 1080p with all maxed, except msaa at 4x.

The chart they show on the first page for 1440p has 970 SLI using about 3.4GB of VRAM. I don't think the drivers start pushing data into the slower segment of memory until it's necessary. The 4k page shows them using about 3.7GB of VRAM.

This is gonna sound silly but what about CF/SLI for older cards like 660's and a 6990's? everywhere I look the oldest card they used was around 750/60.

Those cards are 3 and 4 years old, respectively. With the real-world testing [H] does I'm sure they don't have the time to spend testing a wide range of hardware from different generations. I'm glad they focus on the latest generation of stuff and get as detailed as they do.

I'm sure this has been asked before but is you test rig ever going to change CPU's. My 970 sli numbers seem to be higher than the articles.

I was thinking the same thing. With DX11 the weakest point is still the CPU, even with the NVIDIA drivers enabling multithreading. SLI adds even more CPU overhead. I think a LGA 2011-v3 processor like the i7-5930k would go a long way in alleviating some of the CPU bottleneck being seen in these mGPU reviews.

 

[Cataulin]

Those cards are 3 and 4 years old, respectively. With the real-world testing [H] does I'm sure they don't have the time to spend testing a wide range of hardware from different generations. I'm glad they focus on the latest generation of stuff and get as detailed as they do.

Im not knocking what Kyle and Steve do but lots of people still use the 6/7XXX and 6/7xx series cards. some are using even older. Its very possible that they dont have any old cards but it wouldnt hurt just to run them and see right?

 

[Bankie]

Im not knocking what Kyle and Steve do but lots of people still use the 6XXX series cards. some are using even older. Its very possible that they dont have any old cards but it wouldnt hurt just to run them and see right?

It would be nice if they could include one older card (say a 660/70 or 7970) in the reviews just to see what an upgrade to the latest card will get you. I currently have a 670 and while I can ballpark it (it should be around 760/770/960 performance most of the time) it doesn't always work out that way.

I know this has been asked multiple times over the years and Kyle or Brent say that it just takes too much time that would be better spent on the latest hardware. But maybe they could include just one card from a couple of generations ago as a control.

 

[Dayaks]

Thanks for the review. I hope [H] will be doing reviews on the Witcher 3... Looking forward to that one. The feature reviews are an especially useful resource.

 

[Elmy]

I am running a Club3D 295X2 on GTA on a Benq 2730Z 1440p 144Hz monitor and I don't see any stuttering at all....

 

[No Nrg]

Is it really highest settings if you don't turn on MSAA or Reflection MSAA?

Was the performance hit too high to include those settings?

Will there be a benchmark done showing the impact of MSAA at various resolutions?

 

[Daroller]

I disagree with the [H] conclusion that FXAA is sufficient in this game. There is far too much aliasing in this game for me to consider it playable at less than 4xMSAA. I've been bouncing back and forth between 8xMSAA and 4xTXAA. They both have their uses, and the performance hit is roughly identical. MSAA is slightly slower, but not noticeable in gameplay.

TITAN X SLI is playable at 1440p with Grass Quality set to Very High, motion blur set to zero, and everything else (including MSAA / reflection MSAA / advanced graphics) set to maximum values. 34fps minimum in the mountains, 63fps average overall. I've beaten the game twice on these settings.

 

[Quartz-1]

Can I pick you up on two points?

Firstly, please make the colours in the graphs consistent. Pick one colour for each card / set of cards and stick with it throughout.

Secondly, on the conclusion page you state that three Titan X cards will be required to max out the game. Yet I couldn't find any evidence in the article to back that up. If I haven't missed the absence, you should either remove that statement or provide evidence to back it up.

 

[lee0539]

how much do you think the reason for scaling issues with titan x is due to drivers?

 

[Flip221]

Wait - Which AMD card was used? Was is R290 or R290x? Shouldn't the R290x have been used if comparing to the GTX980?

 

[Yakk]

I am running a Club3D 295X2 on GTA on a Benq 2730Z 1440p 144Hz monitor and I don't see any stuttering at all....

I wonder if HT could be affecting results?

 

[Brent_Justice]

Aha, on 1440p page, it stated that the highest playable settings for each setup, all of them uses FXAA.

Might account for the lower VRAM usage

EDIT:

Kyle, any chance you can clarify what kind of AA, if any, were used in your performance tests?

The review stated FXAA in the beginning of each resolution, but rest of the article is completely silent on the matter.

FXAA was used in all the tests, if it says FXAA on the table, then FXAA is the AA method used.

FXAA iS an AA method, shader based method, doesn't use VRAM, very fast, and reduces jaggies on all polygon edges, textures, alpha textures etc... it is a full-scene algorithm. MSAA, doesn't AA everything (only polygon edges), takes a large performance hit, and eats up precious VRAM that could be used to turn on other in-game features first like high resolution shadows and other features that demand that finite VRAM.

 

[Quartz-1]

Just to pick up on FXAA, does it really produce more of a quality improvement than other features with a similar performance hit? I'm particularly thinking of higher-dpi monitors.

 

[Brent_Justice]

Great review. I wonder what 2 295x2s in 4k maxed out would pull? Vram usage, and power usage would be good to know as well.

Without testing such a scenario, my initial thought would be that scaling would show dimensioning returns. Scaling up to 4 GPUs has never been perfect, even at the best of times. With the game only doing about 60% scaling on 2 GPUs, 4 would be even worse.

It would also potentially create more of the stuttering I experienced scaling up that much with the current state of the profile.

As for VRAM, no doubt the 4GB VRAM would be maxed, I wouldn't recommend doing 4 GPUs unless you had 6-8GB GPUs.

 

[Brent_Justice]

if the 970 only has 3.5 gigs of vram, how is it possible to play at max settings on 1440p?

I use 290s in crossfire, and it uses 3.6 gigs vram at 1080p with all maxed, except msaa at 4x.

Because the game doesn't exceed 4GB at 1440p. It is just enough VRAM to keep the game from stutterying at 1440p. Any higher, and it would.

 

[Brent_Justice]

That is probably the most misconcepted fact about 970...

It has 4GB VRAM, segmented sure, but it still has 4GB.

Also it is based on different architecture, it might have used VRAM a little differently.

Every game so far I've tested never have been able to get the VRAM anywhere near 4GB of usage. This game has shown the highest VRAM I've ever experienced on it so far at 3.7 GB.

 

[Brent_Justice]

I thought they may have used 2xmsaa for the baseline "max settings".

No, FXAA

Priority was put on in-game quality settings, over AA, FXAA is sufficient for reducing AA. I'd rather "max out" the in-game settings, before MSAA.

 

[Brent_Justice]

I'm trying to figure out what you did different in the 4K section. I have 2 Titan Xs in SLI and running the game with everything maxed out except MSAA is off and Grass is on High. The game shows I'm using 8054 out of 24547? If I turn on MSAA and Grass to Ultra I'm up to almost 10,000 used?

Should this number be divided by 1/2 (1/2 per card)?

Thanks!

I don't know, I didn't do anything special, what you see is what it is.

I would say the VRAM counter the game uses in-game doesn't necessarily match with real-world VRAM usage. We used an external third party program to monitor VRAM usage on each GPU.

 

[madgun]

I have a 4K monitor and see the game consuming close to 6.2 Gigs of VRAM on my Titan X. This game can be a VRAM hog if you turn the settings up. On everything set to the highest settings with High-Res shadows off, I get a consistent 60 fps on my Titan X with some scenes dropping to 40. Overall I am extremely please with my Titan X purchase.

 

[Brent_Justice]

Well they say at the start, that any mention of the "maxed out" settings will use exactly what the screen shots show, so it has to be FXAA.

I didn't think alot of people would be using FXAA due to the blurriness it creates, though i may be wrong i guess, but i am not sure saying that 980 SLI is great for 2560x1440p at "maxed" settings is completely correct, change the AA to anything but FXAA and you will more than likely go over the VRAm buffer at the mentioned settings.

I am seeing 4.5gb+ VRAM usage with the same setiings using 2x MSAA(MFAA enabled) on my Titan X




Yeah, you have to halve that, best to use a program like MSI Afterburner or something to run an on screen display while actually playing the game to see you VRAM usage anyway.

FXAA was used.

FXAA allows the removal of jaggies on polygon edges, alpha textures and other textures in the game. MSAA is limited to just the removal of jaggies on polygon edges.

FXAA is the equivelent of 4X MSAA, at no or little performance cost. FXAA also does not eat into the previous finite VRAM.

FXAA is the best option if you want to maximize the in-game quality settings. We put priority first on enabling all the in-game quality options the developer put in the game. Once we have "maxed out" that, using FXAA to reduce AA, then we would start to see if any MSAA levels are playable. But you are still losing things that MSAA reduces jaggies on. FXAA does a much superior job of removing jaggies on EVERYTHING. In a game like this, where there are long view distances, and many types of textures that are not physical objects with lines, FXAA can do a lot to improve image quality.

 

[Brent_Justice]

I'm with you on that, MSAA is the real VRAM killer in this game, as soon as you turn it on VRAM usage skyrockets.

With all the intricate detail in GTA V, FXAA really hurts image quality IMO - can't stand the blur.

If our goal was to just push VRAM up past its limits, sure, go MSAA. But that makes no sense. There is a finite amout of VRAM to use on each video card. We have to work within our limits. FXAA allows us to work within those limits, and turn on other in-game settings that benefit image quality and improve the gameplay experience.

For example, if I had the choice: High Resolution Shaodws + FXAA, or Low Resolution Shadows + 2X MSAA. I'm going with option 1.

FXAA allows us to turn on other features that also need that precious VRAM, but also improve the visual experience more.

 

[Brent_Justice]

Is it really highest settings if you don't turn on MSAA or Reflection MSAA?

Was the performance hit too high to include those settings?

Will there be a benchmark done showing the impact of MSAA at various resolutions?

Yes, MSAA is just one form of AA available in the game. FXAA, and TXAA are two other forms of AA. MSAA is not superior to FXAA in terms of what it removes jaggies on. MSAA can only remove jaggies on polygon edges, objects. It lacks the ability to remove jaggies on everything else. FXAA can, it reduces jaggies on polygon edges, textures, alpha textures, everything, doesn't matter what because it is a full-scene method. They work very differently.

We put priority on enabling all the in-game quality settings the developer has put into the game first. Shadows, Tessellation, Ambient Occlusion, Grass, Textures, Object Detail, View Distances, Shadow Distances, all of these features are what we want to "maximize" first. AA is secondary to that. Adjust for in-game settings first, then AA.

FXAA allows us to use those in-game settings at higher levels than MSAA would. MSAA would both A.) reduce the in-game settings we could use and B.) eat away at VRAM also reducing what we can use and C.) reduce performance a great deal more.

In a game that is already VRAM sensitive, already slow performing, already demanding, and needs the visual improvement of reducing AA on everything, MSAA doesn't make a lot of sense. UNLESS, you just have performance to spare, AND WE DON'T.

Will I test MSAA performance and image quality? Heck yeah! Look for it in part 3.

 

[Brent_Justice]

I disagree with the [H] conclusion that FXAA is sufficient in this game. There is far too much aliasing in this game for me to consider it playable at less than 4xMSAA. I've been bouncing back and forth between 8xMSAA and 4xTXAA. They both have their uses, and the performance hit is roughly identical. MSAA is slightly slower, but not noticeable in gameplay.

TITAN X SLI is playable at 1440p with Grass Quality set to Very High, motion blur set to zero, and everything else (including MSAA / reflection MSAA) set to maximum values. 34fps minimum in the mountains, 63fps average overall. I've beaten the game twice on these settings.

You are welcome to disagree.

Factually though, FXAA reduces aliasing on more things than MSAA does. FXAA reduces aliasing on polygon edges (objects) like MSAA, but also extends that to textures and alpha textures (trees, foliage, vegetation, fences, grass) any form of texture. FXAA is able to work inside polygons, where MSAA cannot. FXAA is a full-scene method. In a game like this, where view distance is important, FXAA does a better job "in the distance."

Technically therefore, FXAA reduces aliasing on more than MSAA does.

We will of course be looking at AA performance and IQ in Part 3.