WEBVTT 00:00:00.280 --> 00:00:02.080 When NVIDIA released SLI, 00:00:02.080 --> 00:00:04.780 it enabled gamers to enjoy next generation levels 00:00:04.780 --> 00:00:06.760 of quality today. 00:00:06.760 --> 00:00:07.640 Assuming of course, 00:00:07.640 --> 00:00:10.000 that they could afford literally twice 00:00:10.000 --> 00:00:11.720 as many graphics cards. 00:00:11.720 --> 00:00:15.000 But as monitor resolutions have grown, 00:00:15.000 --> 00:00:18.080 the scalable part of the scalable link interface, 00:00:18.080 --> 00:00:20.280 which has been with us for over a decade, 00:00:20.280 --> 00:00:23.600 hasn't been able to keep up in spite of fancy, 00:00:23.600 --> 00:00:26.580 high bandwidth bridges like this one. 00:00:26.580 --> 00:00:29.900 Meanwhile, over on the professional side of things, 00:00:29.900 --> 00:00:30.980 NVIDIA has been pushing 00:00:30.980 --> 00:00:35.980 a newer inter-GPU communication protocol called NVLink. 00:00:38.350 --> 00:00:41.650 This essentially turns SLI up to 11. 00:00:41.650 --> 00:00:46.940 But why would you, the general consumer, care about that? 00:00:46.980 --> 00:00:51.560 Well, because NVLink is coming to consumers 00:00:51.560 --> 00:00:54.340 with the GeForce RTX series. 00:00:54.340 --> 00:00:58.600 So it is time then to ask the big question. 00:00:58.600 --> 00:01:00.960 Does it make gaming 00:01:00.960 --> 00:01:01.800 better? 00:01:01.800 --> 00:01:06.090 Whew, that's a lot of hardware. 00:01:06.090 --> 00:01:07.530 Speaking of big questions, 00:01:07.530 --> 00:01:09.050 have you tried GlassWire? 00:01:09.050 --> 00:01:12.150 Detect malware and block badly behaving apps 00:01:12.150 --> 00:01:14.010 on your PC or Android device. 00:01:14.010 --> 00:01:17.870 Use offer code Linus to get 25% off GlassWire 2.0 00:01:17.870 --> 00:01:28.240 at the link below. 00:01:28.240 --> 00:01:29.780 So one of the first things you'll notice 00:01:29.780 --> 00:01:31.940 about a card equipped with NVLink 00:01:31.940 --> 00:01:35.500 is just how big the connector fingers are 00:01:35.500 --> 00:01:37.240 compared to traditional SLI. 00:01:37.240 --> 00:01:40.340 They are more than three times as wide with way, 00:01:40.340 --> 00:01:41.900 way more pins. 00:01:41.900 --> 00:01:45.440 Like seriously, a single NVLink finger 00:01:45.440 --> 00:01:49.460 is wider than the entire SLI connector setup. 00:01:49.460 --> 00:01:52.900 It almost even looks like they're little 00:01:52.900 --> 00:01:55.580 PCI Express connectors, which, 00:01:55.580 --> 00:01:58.860 as we're about to see, isn't by accident. 00:01:58.860 --> 00:02:03.620 So the way that SLI works is actually a lot like, 00:02:03.620 --> 00:02:06.060 oh, here, I have a good prep for this. 00:02:06.060 --> 00:02:09.460 It's actually a lot like the older SCSI and IDE. 00:02:09.460 --> 00:02:14.280 One card functions as the master in the relationship 00:02:14.280 --> 00:02:16.580 and the other one as a slave, 00:02:16.580 --> 00:02:18.740 or in the case of multiple other cards, 00:02:18.740 --> 00:02:20.860 they would all then be slaves. 00:02:20.860 --> 00:02:24.420 So that means that because the master alone 00:02:24.420 --> 00:02:27.140 is directing the workload for those slave cards 00:02:27.140 --> 00:02:31.520 with at best two gigabytes per second of bandwidth 00:02:31.520 --> 00:02:34.300 using one of NVIDIA's high bandwidth bridges, 00:02:34.300 --> 00:02:36.940 you've got enough for the render results 00:02:36.940 --> 00:02:38.820 to be returned to the master 00:02:38.820 --> 00:02:43.080 and honestly, not a whole lot more. 00:02:43.080 --> 00:02:47.040 This is the reason why you can't simply add together 00:02:47.040 --> 00:02:49.540 the memory of your SLI graphics cards, 00:02:49.540 --> 00:02:51.520 taking two 11 gig cards and saying, 00:02:51.520 --> 00:02:54.620 well, I've got 22 gigs of RAM now. 00:02:54.620 --> 00:02:57.840 And the same is true for Team Red's Crossfire. 00:02:57.840 --> 00:03:02.780 By contrast, NVLink is bi-directional 00:03:02.780 --> 00:03:05.780 and it's configured as a mesh, 00:03:05.780 --> 00:03:08.900 which means that no one card is the master, 00:03:08.900 --> 00:03:11.420 and there are no slaves. 00:03:11.420 --> 00:03:14.980 Think of it more like if you were plugging computers 00:03:14.980 --> 00:03:18.240 into a router or a switch. 00:03:18.240 --> 00:03:21.540 So this, along with the extra pins 00:03:21.540 --> 00:03:23.820 and newer signaling protocol, 00:03:23.820 --> 00:03:27.380 gives these cards a lot more bandwidth, 00:03:27.380 --> 00:03:29.440 more than even PCI Express, 00:03:29.440 --> 00:03:34.440 at a total of up to 160 to 300 gigabytes per second. 00:03:35.620 --> 00:03:37.360 That kind of speed lets them pool resources, 00:03:37.360 --> 00:03:38.360 that kind of speed lets them pool resources, 00:03:38.360 --> 00:03:39.060 that kind of speed lets them pool resources, 00:03:39.060 --> 00:03:43.080 in a way that allows access to each card's memory 00:03:43.080 --> 00:03:46.740 and CUDA cores as though they were a single card. 00:03:46.740 --> 00:03:49.220 And that's perfect for the scientific 00:03:49.220 --> 00:03:51.240 and high-end render stations 00:03:51.240 --> 00:03:55.480 that NVIDIA has traditionally targeted with NVLink. 00:03:55.480 --> 00:03:58.040 Now, you might be thinking to yourself, 00:03:58.040 --> 00:04:02.160 awesome, NVLink is coming to GeForce RTX cards. 00:04:02.160 --> 00:04:03.640 We're gonna get those benefits. 00:04:03.640 --> 00:04:06.190 I'm doubling my pre-order. 00:04:06.190 --> 00:04:08.490 Hold your horses there, Tom. 00:04:08.490 --> 00:04:09.170 Yeah, it's awesome. 00:04:09.170 --> 00:04:12.030 But the number of links provided on RTX 00:04:12.030 --> 00:04:13.450 is relatively minimal, 00:04:13.450 --> 00:04:18.450 and the RTX cards only support SLI over the NVLink bus. 00:04:19.610 --> 00:04:24.010 So there will be no fancy resource pooling going on here. 00:04:24.010 --> 00:04:28.750 So our plan today then is to take our Quadro GP100s 00:04:28.750 --> 00:04:31.150 and run them both in compute mode, 00:04:31.150 --> 00:04:33.450 which actually disables the graphics engine, 00:04:33.450 --> 00:04:34.810 like we couldn't plug a display 00:04:34.810 --> 00:04:36.730 into these things right now if we tried, 00:04:36.730 --> 00:04:38.470 and in what's called SLI mode, 00:04:38.470 --> 00:04:41.350 to look at their gaming performance. 00:04:41.350 --> 00:04:43.070 Yes, yes, I know. 00:04:43.070 --> 00:04:45.750 This card isn't intended for gaming, 00:04:45.750 --> 00:04:49.730 but if you look closely at the spec of it, 00:04:49.730 --> 00:04:53.610 it's got HBM2 memory, yes, and more of it, 00:04:53.610 --> 00:04:58.610 but it's otherwise actually very similar to the GTX 1080 Ti. 00:04:59.750 --> 00:05:04.050 So this is probably as close as we will ever get 00:05:04.050 --> 00:05:08.050 to an apples-to-apples comparison between SLI and NVLink. 00:05:08.470 --> 00:05:13.270 Since Pascal is likely to be the only generation of products 00:05:13.270 --> 00:05:16.530 where both of these technologies are present. 00:05:16.530 --> 00:05:18.010 First up, some pre-flight tweaks 00:05:18.010 --> 00:05:19.550 to get everything working though. 00:05:19.550 --> 00:05:22.370 We needed a Quadro SLI-certified motherboard, 00:05:22.370 --> 00:05:26.170 so our ASUS X299 Deluxe with a Core i9-7900X 00:05:26.170 --> 00:05:27.630 worked nicely for this. 00:05:27.630 --> 00:05:31.450 And to look at NVLink's non-gaming performance, 00:05:31.450 --> 00:05:33.310 we needed to configure both cards 00:05:33.310 --> 00:05:35.750 in Tesla compute cluster mode, 00:05:35.750 --> 00:05:38.390 which we can check by going ahead and running... 00:05:38.470 --> 00:05:41.370 this command in the Windows PowerShell. 00:05:41.370 --> 00:05:42.790 So you can see right here, 00:05:42.790 --> 00:05:46.230 links one to three, or zero to three, excuse me, 00:05:46.230 --> 00:05:48.090 or one to, whatever the point is, 00:05:48.090 --> 00:05:52.610 they're all running, and that's good. 00:05:52.610 --> 00:05:55.930 Unfortunately, many of our benchmarks 00:05:55.930 --> 00:05:59.650 actually didn't cooperate very well 00:05:59.650 --> 00:06:03.740 with this particular setup, 00:06:03.740 --> 00:06:07.580 though the latest experimental Blender build managed it, 00:06:07.580 --> 00:06:08.420 and, whew! 00:06:09.360 --> 00:06:14.110 The results pretty much speak for themselves. 00:06:14.110 --> 00:06:17.200 Three and a half minutes for Gooseberry? 00:06:17.200 --> 00:06:20.700 20 seconds for BMW? 00:06:20.700 --> 00:06:22.300 In spite of these tests 00:06:22.300 --> 00:06:24.960 not being particularly memory intensive, 00:06:24.960 --> 00:06:28.180 we are seeing a clear advantage here. 00:06:28.180 --> 00:06:31.660 As for Luxmark's lower OpenCL performance scaling, 00:06:31.660 --> 00:06:35.320 that suggests that CUDA is a necessary ingredient 00:06:35.320 --> 00:06:38.260 if we wanna take full advantage of NVLink. 00:06:38.260 --> 00:06:39.800 Big surprise, of course. 00:06:40.100 --> 00:06:41.480 Not all there is to it, though. 00:06:41.480 --> 00:06:43.900 Remember how NVLink allows us to utilize 00:06:43.900 --> 00:06:45.960 all of the available memory on our cards 00:06:45.960 --> 00:06:48.470 as though they were one big card? 00:06:48.470 --> 00:06:50.310 Well, because of that, 00:06:50.310 --> 00:06:52.870 we can now work with much larger data sets 00:06:52.870 --> 00:06:55.930 than would have been possible on smaller configurations. 00:06:55.930 --> 00:06:58.610 And trust us, we tried on smaller configurations. 00:06:58.610 --> 00:07:02.610 You can see here, even our twinned GP100s 00:07:02.610 --> 00:07:05.130 couldn't handle this particular workload. 00:07:05.130 --> 00:07:09.870 So, it's time to bring out the big guns. 00:07:09.870 --> 00:07:14.410 Our GV100s with their new NVLink bridges 00:07:14.410 --> 00:07:19.410 will give us a total of 64 gigs of HBM2 memory. 00:07:19.710 --> 00:07:21.750 That's more than the system memory 00:07:21.750 --> 00:07:24.170 of even many workstations. 00:07:24.170 --> 00:07:25.810 And there it is. 00:07:25.810 --> 00:07:29.630 Our GV100s handle this just fine. 00:07:29.630 --> 00:07:33.910 So, that's super impressive and extremely useful 00:07:33.910 --> 00:07:37.630 for people with huge data sets. 00:07:37.630 --> 00:07:39.410 But the real thing we were after here was 00:07:39.410 --> 00:07:44.410 evaluating the SLI mode that is coming with the RTX series. 00:07:44.570 --> 00:07:48.170 So, and here we go. 00:07:48.170 --> 00:07:50.990 So, in a massive surprise to no one, 00:07:50.990 --> 00:07:55.990 the GV100s are the fastest solution on the block, for now. 00:07:56.410 --> 00:07:59.590 In SLI, even at 4K Ultra, 00:07:59.590 --> 00:08:04.030 the average frame rate never dipped below 60, which is huge. 00:08:04.030 --> 00:08:07.510 Nothing else can even come close to claiming that. 00:08:07.510 --> 00:08:09.250 What's more interesting, though, 00:08:09.250 --> 00:08:12.950 is when we look at the scaling figures side by side. 00:08:12.950 --> 00:08:15.910 So, our GP100s here, 00:08:15.910 --> 00:08:20.750 these guys seem to scale better than the GV100s 00:08:20.750 --> 00:08:22.290 in gaming and productivity, 00:08:22.290 --> 00:08:24.590 giving them the best scaling overall, 00:08:24.590 --> 00:08:28.210 which may suggest some kind of CPU bottleneck. 00:08:28.210 --> 00:08:30.510 As for the GTX 1080 Ti, 00:08:30.510 --> 00:08:33.530 well, there's huge gains to be made in gaming, 00:08:33.530 --> 00:08:35.850 but not as much in productivity. 00:08:35.850 --> 00:08:38.570 So, as you might expect with anything new, 00:08:39.250 --> 00:08:42.870 I think SLI doesn't scale the same way for everything, 00:08:42.870 --> 00:08:46.490 but it does look to be a pretty decent improvement 00:08:46.490 --> 00:08:48.210 over traditional SLI, 00:08:48.210 --> 00:08:51.750 about 10 to 23% better by our measure, 00:08:51.750 --> 00:08:53.930 with the potential to dramatically improve 00:08:53.930 --> 00:08:56.970 undesirable behavior like micro-stuttering as well, 00:08:56.970 --> 00:08:59.970 or even enable more than two-way SLI 00:08:59.970 --> 00:09:02.490 with decent scaling in the future. 00:09:02.490 --> 00:09:05.970 That is, depending on how much NVIDIA decides to neuter it 00:09:05.970 --> 00:09:08.650 compared to its professional-grade cousin. 00:09:08.650 --> 00:09:09.010 You, too, never know. 00:09:09.010 --> 00:09:10.070 You never know with those guys. 00:09:10.070 --> 00:09:11.990 I mean, one thing we discovered 00:09:11.990 --> 00:09:13.650 in the course of our testing for this video 00:09:13.650 --> 00:09:16.410 is that the new NVLink bridges here 00:09:16.410 --> 00:09:20.770 don't work with the old NVLink cards, even the pro ones. 00:09:20.770 --> 00:09:23.310 So, NVIDIA told us something about 00:09:23.310 --> 00:09:26.330 consumer NVLink bridges having fewer pins, 00:09:26.330 --> 00:09:29.050 or more importantly, a slightly different pin-out, 00:09:29.050 --> 00:09:30.590 but could they have made it work? 00:09:30.590 --> 00:09:31.430 I don't know. 00:09:31.430 --> 00:09:32.330 I don't know with those guys. 00:09:32.330 --> 00:09:36.150 Either way, NVLink has lots of potential 00:09:36.150 --> 00:09:38.310 and looks like a significant hardware upgrade 00:09:38.310 --> 00:09:39.550 that should only improve 00:09:39.550 --> 00:09:42.330 as the drivers themselves continue to improve. 00:09:42.330 --> 00:09:46.650 So maybe, just maybe, SLI isn't dead. 00:09:46.650 --> 00:09:48.450 Yet. 00:09:48.450 --> 00:09:50.250 Maybe. 00:09:50.250 --> 00:09:51.310 But you know what's not maybe? 00:09:51.310 --> 00:09:54.330 FreshBooks, being the small business accounting software 00:09:54.330 --> 00:09:56.510 custom-built for how you wanna work. 00:09:56.510 --> 00:09:58.350 If you're a freelancer or a small business owner, 00:09:58.350 --> 00:10:00.110 you need to check out FreshBooks. 00:10:00.110 --> 00:10:01.850 It's the simple way to be more productive, 00:10:01.850 --> 00:10:04.310 more organized, and to get paid faster. 00:10:04.310 --> 00:10:06.610 You can create and send professional-looking invoices 00:10:06.610 --> 00:10:08.070 in less than 30 seconds. 00:10:08.070 --> 00:10:09.410 You can set up online payments, 00:10:09.410 --> 00:10:10.570 with just a couple of clicks, 00:10:10.570 --> 00:10:12.470 to get paid up to four days faster, 00:10:12.470 --> 00:10:14.910 and you don't have to take my word for it. 00:10:14.910 --> 00:10:17.290 Go try FreshBooks for 30 days for free 00:10:17.290 --> 00:10:19.450 at freshbooks.com slash techtips. 00:10:19.450 --> 00:10:20.770 Just enter Linus Tech Tips 00:10:20.770 --> 00:10:22.270 in the How You Heard About Us section, 00:10:22.270 --> 00:10:24.450 so they'll know who sent you. 00:10:24.450 --> 00:10:25.430 So thanks for watching, guys. 00:10:25.430 --> 00:10:27.230 If this video sucked, you know what to do. 00:10:27.230 --> 00:10:28.910 But if it was awesome, get subscribed, 00:10:28.910 --> 00:10:32.330 hit that like button, or check out the NVLink. 00:10:32.330 --> 00:10:33.990 Oh, Lordy, that's awful. 00:10:33.990 --> 00:10:35.910 Anthony, come on. 00:10:35.910 --> 00:10:37.250 To where to buy the stuff we featured 00:10:37.250 --> 00:10:38.330 in the video description. 00:10:38.330 --> 00:10:39.250 Also, NVLink. 00:10:39.410 --> 00:10:40.250 NVLinked. 00:10:40.250 --> 00:10:41.450 In the description is our merch store, 00:10:41.450 --> 00:10:43.870 which has cool shirts, and our community forum, 00:10:43.870 --> 00:10:45.370 which you should totally join. 00:10:48.540 --> 00:10:50.300 She made me do it twice. 00:10:50.300 --> 00:10:52.420 I mean, it's really, it's my fault. 00:10:52.420 --> 00:10:54.420 I'll read anything on the teleprompter.