MultiGPU scaling tests by DG Lee.

cageymaru

cageymaru

Fully [H]
2FA
Joined
Apr 10, 2003
Messages
23,539
DG Lee conducted MultiGPU tests on the FURY X, 980ti, and Titan X. It was quite interesting to see how they scaled from 1 card to 2, then 3, and finally quad cards.
http://iyd.kr/753

Here is an example. There are plenty of tests conducted. I remember someone on these boards Bashaa? who ran into the same issue months ago with quad Titan X. Anyways I saw this on WCCFTECH initially.

RNJ7Fj6.png


Setting used in the testing. Thanks to Derangel for linking!

sPRHG1q.png
 
Last edited:
Certainly very interesting. One thing I am curious about though is how every series of benchmarks I've seen so far have shown these scaling issues at 100+ fps. Seems like most 4k displays are lucky to even do 60hz. Could this simply be something that was taken into account in the drivers?

I'd love to see scaling tests done with something that could actually bog it down to <60fps even with several cards, if something of that nature even exists.
 
Too bad they didn't include what settings were used. It seems though that dual crossfire is worth it @ UHD. Finding another Fury X, however, is almost impossible right now.
 
Increase in performance expressed in percentage over single card instead of flat FPS would have made the analysis easier.
Edit: Ignore this as relative performance is discussed later in the article.
 
Do these number represent real gameplay or just useless canned benchmarks?:rolleyes:
 
I like how the one "example" he posts is the only one where the Fury X leads by any perceptible margin. Certainly no agenda here.
 
LigTasm said:
I like how the one "example" he posts is the only one where the Fury X leads by any perceptible margin. Certainly no agenda here.
Click to expand...

Actually the war i read it, all the examples come to same conclusion. And that is the fury x looses almost every time except in multi gpu and the more gpus added the better. It also seems to shine the best in UHD res somehow while having the least amount of vram. But like others have said it doesn't actually show how playable any of the games were. Pretty graph but not overly useful to me.
 
Alien Isolation is very playable on a single 290x in UHD. So not much point to this exercise.
 
LigTasm said:
I like how the one "example" he posts is the only one where the Fury X leads by any perceptible margin. Certainly no agenda here.
Click to expand...

Oh I wanted to show the biggest difference. If anything I want Nvidia users to go back to Nvidia and demand that they fix their drivers. Nobody purchasing from AMD or Nvidia and spending thousands of dollars should experience bad support for their purchases like the example in the OP.

In short yes, it was meant to send a message. Now go complain to them!

(Now watch the Nvidia crew come out and try and save face so that nobody will complain and Nvidia doesn't feel compelled to fix it.)
 
M76 said:
Alien Isolation is very playable on a single 290x in UHD. So not much point to this exercise.
Click to expand...

There were many games tested. You should go and at least look at the charts. It is worthwhile.
 
cageymaru said:
There were many games tested. You should go and at least look at the charts. It is worthwhile.
Click to expand...

I meant showing this chart as an example is pointless since the game can be played on a single gpu in uhd.

Anyway all I can take away from the charts is that Quad SLI or Quadfire is kind of pointless. Either it scales badly or the game already has enough fps with less gpus.
 
M76 said:
I meant showing this chart as an example is pointless since the game can be played on a single gpu in uhd.

Anyway all I can take away from the charts is that Quad SLI or Quadfire is kind of pointless. Either it scales badly or the game already has enough fps with less gpus.
Click to expand...

I'd say everyone buying three or four cards for a rig understands that. For enthusiasts with the money to burn it's more of a "it looks cool" or "why the hell not" factor. By the time you add even a second one of these cards you have long ago abandoned the idea of performance per dollar.
 
M76 said:
I meant showing this chart as an example is pointless since the game can be played on a single gpu in uhd.

Anyway all I can take away from the charts is that Quad SLI or Quadfire is kind of pointless. Either it scales badly or the game already has enough fps with less gpus.
Click to expand...

Let's not forget UHD Eyefinity/Surround. 3 UHD monitors need more than 1 card. But you're right, if you have one UHD display and can push 60Hz there's no point.
 
M76 said:
I meant showing this chart as an example is pointless since the game can be played on a single gpu in uhd.

Anyway all I can take away from the charts is that Quad SLI or Quadfire is kind of pointless. Either it scales badly or the game already has enough fps with less gpus.
Click to expand...

At this time you're probably right. But at 4K resolutions where these GPUs were designed to shine it matters. Virtual Reality headsets that are launching this year and next are first generation and will be gimped. Reading countless VR promotional literature it seems that the big players want a product that is 4K per eye and 120Hz minimum and less than 5ms of frame time in the future.

Currently the Oculus Rift is 2160x1200 due to a lack of screen technology. Samsung is developing 11k screens for debut at the Olympics in 2018; think Samsung Gear VR headsets. So getting AMD and Nvidia on board for launches like this is important.

I think it's fine to expect more from them. It gives the engineers something to aspire to.
 
cageymaru said:
Oh I wanted to show the biggest difference. If anything I want Nvidia users to go back to Nvidia and demand that they fix their drivers. Nobody purchasing from AMD or Nvidia and spending thousands of dollars should experience bad support for their purchases like the example in the OP.

In short yes, it was meant to send a message. Now go complain to them!
Click to expand...

I don't even have an Nvidia card. In this test, both companies scaled like shit past 2 cards. Who cares, nothing is new there.
 
LigTasm said:
I don't even have an Nvidia card. In this test, both companies scaled like shit past 2 cards. Who cares, nothing is new there.
Click to expand...

I know that. But it does matter when it comes to VR unless you like motion sickness. Got to think ahead of today and about tomorrow. ;)
 
I wonder how if NVIDIA is running up against the bandwidth limitations of the SLI bridge, like AMD was before they added the XDMA engine for CrossFire in Hawaii.
 
I though it was well known that SLI technology has not been updated in a while and was due for a revamp.
Pascal?
 
cageymaru said:
I know that. But it does matter when it comes to VR
Click to expand...
Multi-GPU scaling in VR is totally different (these charts aren't relevant in the slightest).

In VR, each GPU becomes dedicated to a different eye, meaning no AFR or SFR issues. Each GPU is free to render its own image as fast as it wants, which SHOULD mean 100% scaling + lower latency than a single card.

Single-GPU workload in VR: Card 1 renders Eye 1 THEN Eye 2.
Multi-GPU workload in VR: Card 1 renders Eye 1 WHILE Card 2 renders Eye 2.
 
Last edited:
You must log in or register to reply here.
Back
Top