Why Don't Multiple GPUs Scale Properly?
Techquickie
·Techquickie
·2017-05-06
·
1,121 words · ~5 min read
0:00
In your quest for the highest frame rates, the prettiest textures, and the
0:05
most karma in the PC master race
0:08
subreddit, you decided to go with a multi-GPU setup for your new gaming rig.
0:14
But when you fire up your games, you notice that the sick gains you're
0:19
getting aren't really what you expected, especially after
0:24
dropping so much money on your system.
0:28
So, what's the deal with that? Well, this problem has plagued PC gamers for
0:33
over a decade. Why don't you usually get
0:37
double the performance from a dual card configuration or triple if you have
0:42
three video cards? I mean, since graphics rendering benefits from having
0:46
more computing units in parallel, which you can learn more about in this video,
0:51
by the way, wouldn't it make sense then for multicard setups like SLI and
0:56
Crossfire to scale pretty easily?
0:59
Unfortunately, it's not quite that simple. For starters, using two separate
1:04
graphics cards means they have to talk to each other somehow. You see, most
1:09
modern multiGPU setups use something called alternate frame rendering, where
1:14
each card renders every other frame
1:17
separately. Sounds efficient, but for
1:20
them to be pushed to your monitor in the right order, the cards have to sync up
1:25
properly, creating a surprising amount of overhead for both the GPUs and the
1:31
electrical communication paths or buses that link them. This overhead not only
1:37
gets worse as you add more cards to your rig, but can even result in a very
1:43
unpleasant phenomenon called micro stuttering, which often shows up even if
1:48
you're getting lots of frames per second. You see, for smooth gameplay,
1:53
you need not only high FPS, but good and
1:58
consistent frame times, which is the
2:02
amount of time that a given frame will remain on your screen. Poor
2:07
synchronization between the cards can leave certain frames on your monitor for
2:12
way longer than they should be, even if the total number of frames each second
2:17
is very high. This can render certain games
2:20
unplayable, even though your setup is more than powerful enough to run it on
2:26
paper. And as if that wasn't enough, an
2:29
even bigger problem is that it's very difficult for your GPUs to know what
2:34
frame to render next since it depends heavily on what the player chooses to
2:40
do. The real time nature of game rendering makes it tough to parallelize
2:46
in certain games among multiple GPUs,
2:49
even with alternate frame rendering, as there isn't always a second frame or
2:54
even part of a predicted one for your second card to work on. Of course, it
3:00
isn't all about the inherent limitations of your hardware. On the software side,
3:05
some game developers have been able to code their games in a manner that makes
3:10
it easier for multiple GPUs to split the workload, while others simply don't
3:15
bother. With the difficulty in coding for a very small subset of customers,
3:22
many devs don't see the point in sinking
3:25
tons of time into making sure that their games will play nice with multicard
3:30
systems. Some games even manage to have worse performance in these situations.
3:36
And finally, you definitely shouldn't run out and buy multiple GPUs if you're
3:40
rocking a multiple monitor setup and you want extra video memory or VRAM to store
3:47
all of your textures. With SLI and
3:50
Crossfire, you are limited only to however much VRAM is on one card because
3:56
both GPUs need to be able to access the same information in memory in order to
4:00
work properly in tandem, whatever it might say on the packaging. So, does
4:06
this mean then that SLI and Crossfire are useless gimmicks and a complete
4:10
waste of money? Well, no. Multicard
4:13
setups will give you the best possible performance in certain use cases and can
4:19
even make some games playable at very high settings and resolutions where they
4:25
wouldn't otherwise be, even with the best single card on the market. Not to
4:29
mention that multiGPU rendering may improve with new APIs like Vulcan and
4:35
DirectX12, which by the way you can learn more about here. Just do your
4:39
research before you buy as to what games
4:42
see the greatest benefit. And remember to keep your expectations realistic.
4:47
After all, you wouldn't expect to be able to drive twice as fast just because
4:52
you bought two cars, would you? Unless
4:55
you're trying to do like a fan remake of Tokyo
5:00
Drift. You guys know that I love Dollar Shave Club. We've been working with them
5:05
for years. Good quality razors, good
5:08
pricing, the shave is fantastic. But what you probably don't know is that
5:13
they have other products, too. Well, though, if you watch our ad
5:17
integrations, we talk a lot about their butt wipes. So, you probably are aware
5:20
of those. But the point is, it's good stuff. They've got their Dr. Carver
5:25
shave butter. They've got their aftershave. And they've got their one
5:28
wipe charlies which make everything from your face all the way to not your face
5:33
look and smell like a million bucks.
5:36
Trust me guys, once you're in the club, you'll see they've got the best grooming
5:40
products and they're affordable and even
5:43
more importantly for nerds like me, it's easy. You don't have to go out of the
5:48
house in order to look like you go out
5:51
of the house. And you can even save money while doing it. I mean, it really
5:55
is mindboggling, isn't it? The economics of being able to afford to have someone
5:59
walk to your door and bring you these for less money than what you would pay
6:03
to go out yourself and get them. I mean, that's how much margin those guys are
6:07
taking from. Anyway, that's besides the point. Right now is your chance to see
6:10
for yourself why so many of us love Dollar Shave Club. If you've never been
6:14
a member before and you've never joined, now's the time. You'll get your first
6:18
month of razors for free. Just pay the shipping and after that, it's only a few
6:22
bucks a month. It's a limited time offer, so join today. Go to
6:26
dollarshaveclub.com/Linus. That's dollarshaveclub.com/Linus. We've got
6:30
that linked in the video description. Thanks for watching, guys. Like if you
6:34
liked, dislike if you dislike. Check out our other channels. Comment with a video suggestion, and
6:38
subscribe. Subscribe. If I wave my arms like this,
6:42
maybe they'll subscribe. Maybe they'll do it this time.