Is NVLink BETTER than SLI??

Linus Tech Tips ·Linus Tech Tips ·2019-05-06 · 1,641 words · ~8 min read
Floatplane YouTube

Transcript

JSON SRT VTT 233
0:00 When NVIDIA released SLI,
0:02 it enabled gamers to enjoy next generation levels
0:04 of quality today.
0:06 Assuming of course,
0:07 that they could afford literally twice
0:10 as many graphics cards.
0:11 But as monitor resolutions have grown,
0:15 the scalable part of the scalable link interface,
0:18 which has been with us for over a decade,
0:20 hasn't been able to keep up in spite of fancy,
0:23 high bandwidth bridges like this one.
0:26 Meanwhile, over on the professional side of things,
0:29 NVIDIA has been pushing
0:30 a newer inter-GPU communication protocol called NVLink.
0:38 This essentially turns SLI up to 11.
0:41 But why would you, the general consumer, care about that?
0:46 Well, because NVLink is coming to consumers
0:51 with the GeForce RTX series.
0:54 So it is time then to ask the big question.
0:58 Does it make gaming
1:00 better?
1:01 Whew, that's a lot of hardware.
1:06 Speaking of big questions,
1:07 have you tried GlassWire?
1:09 Detect malware and block badly behaving apps
1:12 on your PC or Android device.
1:14 Use offer code Linus to get 25% off GlassWire 2.0
1:17 at the link below.
1:28 So one of the first things you'll notice
1:29 about a card equipped with NVLink
1:31 is just how big the connector fingers are
1:35 compared to traditional SLI.
1:37 They are more than three times as wide with way,
1:40 way more pins.
1:41 Like seriously, a single NVLink finger
1:45 is wider than the entire SLI connector setup.
1:49 It almost even looks like they're little
1:52 PCI Express connectors, which,
1:55 as we're about to see, isn't by accident.
1:58 So the way that SLI works is actually a lot like,
2:03 oh, here, I have a good prep for this.
2:06 It's actually a lot like the older SCSI and IDE.
2:09 One card functions as the master in the relationship
2:14 and the other one as a slave,
2:16 or in the case of multiple other cards,
2:18 they would all then be slaves.
2:20 So that means that because the master alone
2:24 is directing the workload for those slave cards
2:27 with at best two gigabytes per second of bandwidth
2:31 using one of NVIDIA's high bandwidth bridges,
2:34 you've got enough for the render results
2:36 to be returned to the master
2:38 and honestly, not a whole lot more.
2:43 This is the reason why you can't simply add together
2:47 the memory of your SLI graphics cards,
2:49 taking two 11 gig cards and saying,
2:51 well, I've got 22 gigs of RAM now.
2:54 And the same is true for Team Red's Crossfire.
2:57 By contrast, NVLink is bi-directional
3:02 and it's configured as a mesh,
3:05 which means that no one card is the master,
3:08 and there are no slaves.
3:11 Think of it more like if you were plugging computers
3:14 into a router or a switch.
3:18 So this, along with the extra pins
3:21 and newer signaling protocol,
3:23 gives these cards a lot more bandwidth,
3:27 more than even PCI Express,
3:29 at a total of up to 160 to 300 gigabytes per second.
3:35 That kind of speed lets them pool resources,
3:37 that kind of speed lets them pool resources,
3:38 that kind of speed lets them pool resources,
3:39 in a way that allows access to each card's memory
3:43 and CUDA cores as though they were a single card.
3:46 And that's perfect for the scientific
3:49 and high-end render stations
3:51 that NVIDIA has traditionally targeted with NVLink.
3:55 Now, you might be thinking to yourself,
3:58 awesome, NVLink is coming to GeForce RTX cards.
4:02 We're gonna get those benefits.
4:03 I'm doubling my pre-order.
4:06 Hold your horses there, Tom.
4:08 Yeah, it's awesome.
4:09 But the number of links provided on RTX
4:12 is relatively minimal,
4:13 and the RTX cards only support SLI over the NVLink bus.
4:19 So there will be no fancy resource pooling going on here.
4:24 So our plan today then is to take our Quadro GP100s
4:28 and run them both in compute mode,
4:31 which actually disables the graphics engine,
4:33 like we couldn't plug a display
4:34 into these things right now if we tried,
4:36 and in what's called SLI mode,
4:38 to look at their gaming performance.
4:41 Yes, yes, I know.
4:43 This card isn't intended for gaming,
4:45 but if you look closely at the spec of it,
4:49 it's got HBM2 memory, yes, and more of it,
4:53 but it's otherwise actually very similar to the GTX 1080 Ti.
4:59 So this is probably as close as we will ever get
5:04 to an apples-to-apples comparison between SLI and NVLink.
5:08 Since Pascal is likely to be the only generation of products
5:13 where both of these technologies are present.
5:16 First up, some pre-flight tweaks
5:18 to get everything working though.
5:19 We needed a Quadro SLI-certified motherboard,
5:22 so our ASUS X299 Deluxe with a Core i9-7900X
5:26 worked nicely for this.
5:27 And to look at NVLink's non-gaming performance,
5:31 we needed to configure both cards
5:33 in Tesla compute cluster mode,
5:35 which we can check by going ahead and running...
5:38 this command in the Windows PowerShell.
5:41 So you can see right here,
5:42 links one to three, or zero to three, excuse me,
5:46 or one to, whatever the point is,
5:48 they're all running, and that's good.
5:52 Unfortunately, many of our benchmarks
5:55 actually didn't cooperate very well
5:59 with this particular setup,
6:03 though the latest experimental Blender build managed it,
6:07 and, whew!
6:09 The results pretty much speak for themselves.
6:14 Three and a half minutes for Gooseberry?
6:17 20 seconds for BMW?
6:20 In spite of these tests
6:22 not being particularly memory intensive,
6:24 we are seeing a clear advantage here.
6:28 As for Luxmark's lower OpenCL performance scaling,
6:31 that suggests that CUDA is a necessary ingredient
6:35 if we wanna take full advantage of NVLink.
6:38 Big surprise, of course.
6:40 Not all there is to it, though.
6:41 Remember how NVLink allows us to utilize
6:43 all of the available memory on our cards
6:45 as though they were one big card?
6:48 Well, because of that,
6:50 we can now work with much larger data sets
6:52 than would have been possible on smaller configurations.
6:55 And trust us, we tried on smaller configurations.
6:58 You can see here, even our twinned GP100s
7:02 couldn't handle this particular workload.
7:05 So, it's time to bring out the big guns.
7:09 Our GV100s with their new NVLink bridges
7:14 will give us a total of 64 gigs of HBM2 memory.
7:19 That's more than the system memory
7:21 of even many workstations.
7:24 And there it is.
7:25 Our GV100s handle this just fine.
7:29 So, that's super impressive and extremely useful
7:33 for people with huge data sets.
7:37 But the real thing we were after here was
7:39 evaluating the SLI mode that is coming with the RTX series.
7:44 So, and here we go.
7:48 So, in a massive surprise to no one,
7:50 the GV100s are the fastest solution on the block, for now.
7:56 In SLI, even at 4K Ultra,
7:59 the average frame rate never dipped below 60, which is huge.
8:04 Nothing else can even come close to claiming that.
8:07 What's more interesting, though,
8:09 is when we look at the scaling figures side by side.
8:12 So, our GP100s here,
8:15 these guys seem to scale better than the GV100s
8:20 in gaming and productivity,
8:22 giving them the best scaling overall,
8:24 which may suggest some kind of CPU bottleneck.
8:28 As for the GTX 1080 Ti,
8:30 well, there's huge gains to be made in gaming,
8:33 but not as much in productivity.
8:35 So, as you might expect with anything new,
8:39 I think SLI doesn't scale the same way for everything,
8:42 but it does look to be a pretty decent improvement
8:46 over traditional SLI,
8:48 about 10 to 23% better by our measure,
8:51 with the potential to dramatically improve
8:53 undesirable behavior like micro-stuttering as well,
8:56 or even enable more than two-way SLI
8:59 with decent scaling in the future.
9:02 That is, depending on how much NVIDIA decides to neuter it
9:05 compared to its professional-grade cousin.
9:08 You, too, never know.
9:09 You never know with those guys.
9:10 I mean, one thing we discovered
9:11 in the course of our testing for this video
9:13 is that the new NVLink bridges here
9:16 don't work with the old NVLink cards, even the pro ones.
9:20 So, NVIDIA told us something about
9:23 consumer NVLink bridges having fewer pins,
9:26 or more importantly, a slightly different pin-out,
9:29 but could they have made it work?
9:30 I don't know.
9:31 I don't know with those guys.
9:32 Either way, NVLink has lots of potential
9:36 and looks like a significant hardware upgrade
9:38 that should only improve
9:39 as the drivers themselves continue to improve.
9:42 So maybe, just maybe, SLI isn't dead.
9:46 Yet.
9:48 Maybe.
9:50 But you know what's not maybe?
9:51 FreshBooks, being the small business accounting software
9:54 custom-built for how you wanna work.
9:56 If you're a freelancer or a small business owner,
9:58 you need to check out FreshBooks.
10:00 It's the simple way to be more productive,
10:01 more organized, and to get paid faster.
10:04 You can create and send professional-looking invoices
10:06 in less than 30 seconds.
10:08 You can set up online payments,
10:09 with just a couple of clicks,
10:10 to get paid up to four days faster,
10:12 and you don't have to take my word for it.
10:14 Go try FreshBooks for 30 days for free
10:17 at freshbooks.com slash techtips.
10:19 Just enter Linus Tech Tips
10:20 in the How You Heard About Us section,
10:22 so they'll know who sent you.
10:24 So thanks for watching, guys.
10:25 If this video sucked, you know what to do.
10:27 But if it was awesome, get subscribed,
10:28 hit that like button, or check out the NVLink.
10:32 Oh, Lordy, that's awful.
10:33 Anthony, come on.
10:35 To where to buy the stuff we featured
10:37 in the video description.
10:38 Also, NVLink.
10:39 NVLinked.
10:40 In the description is our merch store,
10:41 which has cool shirts, and our community forum,
10:43 which you should totally join.
10:48 She made me do it twice.
10:50 I mean, it's really, it's my fault.
10:52 I'll read anything on the teleprompter.