{"video_id":"-e4BvZz6H8Q","title":"Is NVLink BETTER than SLI??","channel":"Linus Tech Tips","show":"Linus Tech Tips","published_at":"2019-05-06T14:53:29Z","duration_s":656,"segments":[{"start_s":0.28,"end_s":2.08,"text":"When NVIDIA released SLI,","speaker":null,"is_sponsor":0},{"start_s":2.08,"end_s":4.78,"text":"it enabled gamers to enjoy next generation levels","speaker":null,"is_sponsor":0},{"start_s":4.78,"end_s":6.76,"text":"of quality today.","speaker":null,"is_sponsor":0},{"start_s":6.76,"end_s":7.64,"text":"Assuming of course,","speaker":null,"is_sponsor":0},{"start_s":7.64,"end_s":10.0,"text":"that they could afford literally twice","speaker":null,"is_sponsor":0},{"start_s":10.0,"end_s":11.72,"text":"as many graphics cards.","speaker":null,"is_sponsor":0},{"start_s":11.72,"end_s":15.0,"text":"But as monitor resolutions have grown,","speaker":null,"is_sponsor":0},{"start_s":15.0,"end_s":18.08,"text":"the scalable part of the scalable link interface,","speaker":null,"is_sponsor":0},{"start_s":18.08,"end_s":20.28,"text":"which has been with us for over a decade,","speaker":null,"is_sponsor":0},{"start_s":20.28,"end_s":23.6,"text":"hasn't been able to keep up in spite of fancy,","speaker":null,"is_sponsor":0},{"start_s":23.6,"end_s":26.58,"text":"high bandwidth bridges like this one.","speaker":null,"is_sponsor":0},{"start_s":26.58,"end_s":29.9,"text":"Meanwhile, over on the professional side of things,","speaker":null,"is_sponsor":0},{"start_s":29.9,"end_s":30.98,"text":"NVIDIA has been pushing","speaker":null,"is_sponsor":0},{"start_s":30.98,"end_s":35.98,"text":"a newer inter-GPU communication protocol called NVLink.","speaker":null,"is_sponsor":0},{"start_s":38.35,"end_s":41.65,"text":"This essentially turns SLI up to 11.","speaker":null,"is_sponsor":0},{"start_s":41.65,"end_s":46.94,"text":"But why would you, the general consumer, care about that?","speaker":null,"is_sponsor":0},{"start_s":46.98,"end_s":51.56,"text":"Well, because NVLink is coming to consumers","speaker":null,"is_sponsor":0},{"start_s":51.56,"end_s":54.34,"text":"with the GeForce RTX series.","speaker":null,"is_sponsor":0},{"start_s":54.34,"end_s":58.6,"text":"So it is time then to ask the big question.","speaker":null,"is_sponsor":0},{"start_s":58.6,"end_s":60.96,"text":"Does it make gaming","speaker":null,"is_sponsor":0},{"start_s":60.96,"end_s":61.8,"text":"better?","speaker":null,"is_sponsor":0},{"start_s":61.8,"end_s":66.09,"text":"Whew, that's a lot of hardware.","speaker":null,"is_sponsor":0},{"start_s":66.09,"end_s":67.53,"text":"Speaking of big questions,","speaker":null,"is_sponsor":0},{"start_s":67.53,"end_s":69.05,"text":"have you tried GlassWire?","speaker":null,"is_sponsor":0},{"start_s":69.05,"end_s":72.15,"text":"Detect malware and block badly behaving apps","speaker":null,"is_sponsor":0},{"start_s":72.15,"end_s":74.01,"text":"on your PC or Android device.","speaker":null,"is_sponsor":0},{"start_s":74.01,"end_s":77.87,"text":"Use offer code Linus to get 25% off GlassWire 2.0","speaker":null,"is_sponsor":0},{"start_s":77.87,"end_s":88.24,"text":"at the link below.","speaker":null,"is_sponsor":0},{"start_s":88.24,"end_s":89.78,"text":"So one of the first things you'll notice","speaker":null,"is_sponsor":0},{"start_s":89.78,"end_s":91.94,"text":"about a card equipped with NVLink","speaker":null,"is_sponsor":0},{"start_s":91.94,"end_s":95.5,"text":"is just how big the connector fingers are","speaker":null,"is_sponsor":0},{"start_s":95.5,"end_s":97.24,"text":"compared to traditional SLI.","speaker":null,"is_sponsor":0},{"start_s":97.24,"end_s":100.34,"text":"They are more than three times as wide with way,","speaker":null,"is_sponsor":0},{"start_s":100.34,"end_s":101.9,"text":"way more pins.","speaker":null,"is_sponsor":0},{"start_s":101.9,"end_s":105.44,"text":"Like seriously, a single NVLink finger","speaker":null,"is_sponsor":0},{"start_s":105.44,"end_s":109.46,"text":"is wider than the entire SLI connector setup.","speaker":null,"is_sponsor":0},{"start_s":109.46,"end_s":112.9,"text":"It almost even looks like they're little","speaker":null,"is_sponsor":0},{"start_s":112.9,"end_s":115.58,"text":"PCI Express connectors, which,","speaker":null,"is_sponsor":0},{"start_s":115.58,"end_s":118.86,"text":"as we're about to see, isn't by accident.","speaker":null,"is_sponsor":0},{"start_s":118.86,"end_s":123.62,"text":"So the way that SLI works is actually a lot like,","speaker":null,"is_sponsor":0},{"start_s":123.62,"end_s":126.06,"text":"oh, here, I have a good prep for this.","speaker":null,"is_sponsor":0},{"start_s":126.06,"end_s":129.46,"text":"It's actually a lot like the older SCSI and IDE.","speaker":null,"is_sponsor":0},{"start_s":129.46,"end_s":134.28,"text":"One card functions as the master in the relationship","speaker":null,"is_sponsor":0},{"start_s":134.28,"end_s":136.58,"text":"and the other one as a slave,","speaker":null,"is_sponsor":0},{"start_s":136.58,"end_s":138.74,"text":"or in the case of multiple other cards,","speaker":null,"is_sponsor":0},{"start_s":138.74,"end_s":140.86,"text":"they would all then be slaves.","speaker":null,"is_sponsor":0},{"start_s":140.86,"end_s":144.42,"text":"So that means that because the master alone","speaker":null,"is_sponsor":0},{"start_s":144.42,"end_s":147.14,"text":"is directing the workload for those slave cards","speaker":null,"is_sponsor":0},{"start_s":147.14,"end_s":151.52,"text":"with at best two gigabytes per second of bandwidth","speaker":null,"is_sponsor":0},{"start_s":151.52,"end_s":154.3,"text":"using one of NVIDIA's high bandwidth bridges,","speaker":null,"is_sponsor":0},{"start_s":154.3,"end_s":156.94,"text":"you've got enough for the render results","speaker":null,"is_sponsor":0},{"start_s":156.94,"end_s":158.82,"text":"to be returned to the master","speaker":null,"is_sponsor":0},{"start_s":158.82,"end_s":163.08,"text":"and honestly, not a whole lot more.","speaker":null,"is_sponsor":0},{"start_s":163.08,"end_s":167.04,"text":"This is the reason why you can't simply add together","speaker":null,"is_sponsor":0},{"start_s":167.04,"end_s":169.54,"text":"the memory of your SLI graphics cards,","speaker":null,"is_sponsor":0},{"start_s":169.54,"end_s":171.52,"text":"taking two 11 gig cards and saying,","speaker":null,"is_sponsor":0},{"start_s":171.52,"end_s":174.62,"text":"well, I've got 22 gigs of RAM now.","speaker":null,"is_sponsor":0},{"start_s":174.62,"end_s":177.84,"text":"And the same is true for Team Red's Crossfire.","speaker":null,"is_sponsor":0},{"start_s":177.84,"end_s":182.78,"text":"By contrast, NVLink is bi-directional","speaker":null,"is_sponsor":0},{"start_s":182.78,"end_s":185.78,"text":"and it's configured as a mesh,","speaker":null,"is_sponsor":0},{"start_s":185.78,"end_s":188.9,"text":"which means that no one card is the master,","speaker":null,"is_sponsor":0},{"start_s":188.9,"end_s":191.42,"text":"and there are no slaves.","speaker":null,"is_sponsor":0},{"start_s":191.42,"end_s":194.98,"text":"Think of it more like if you were plugging computers","speaker":null,"is_sponsor":0},{"start_s":194.98,"end_s":198.24,"text":"into a router or a switch.","speaker":null,"is_sponsor":0},{"start_s":198.24,"end_s":201.54,"text":"So this, along with the extra pins","speaker":null,"is_sponsor":0},{"start_s":201.54,"end_s":203.82,"text":"and newer signaling protocol,","speaker":null,"is_sponsor":0},{"start_s":203.82,"end_s":207.38,"text":"gives these cards a lot more bandwidth,","speaker":null,"is_sponsor":0},{"start_s":207.38,"end_s":209.44,"text":"more than even PCI Express,","speaker":null,"is_sponsor":0},{"start_s":209.44,"end_s":214.44,"text":"at a total of up to 160 to 300 gigabytes per second.","speaker":null,"is_sponsor":0},{"start_s":215.62,"end_s":217.36,"text":"That kind of speed lets them pool resources,","speaker":null,"is_sponsor":0},{"start_s":217.36,"end_s":218.36,"text":"that kind of speed lets them pool resources,","speaker":null,"is_sponsor":0},{"start_s":218.36,"end_s":219.06,"text":"that kind of speed lets them pool resources,","speaker":null,"is_sponsor":0},{"start_s":219.06,"end_s":223.08,"text":"in a way that allows access to each card's memory","speaker":null,"is_sponsor":0},{"start_s":223.08,"end_s":226.74,"text":"and CUDA cores as though they were a single card.","speaker":null,"is_sponsor":0},{"start_s":226.74,"end_s":229.22,"text":"And that's perfect for the scientific","speaker":null,"is_sponsor":0},{"start_s":229.22,"end_s":231.24,"text":"and high-end render stations","speaker":null,"is_sponsor":0},{"start_s":231.24,"end_s":235.48,"text":"that NVIDIA has traditionally targeted with NVLink.","speaker":null,"is_sponsor":0},{"start_s":235.48,"end_s":238.04,"text":"Now, you might be thinking to yourself,","speaker":null,"is_sponsor":0},{"start_s":238.04,"end_s":242.16,"text":"awesome, NVLink is coming to GeForce RTX cards.","speaker":null,"is_sponsor":0},{"start_s":242.16,"end_s":243.64,"text":"We're gonna get those benefits.","speaker":null,"is_sponsor":0},{"start_s":243.64,"end_s":246.19,"text":"I'm doubling my pre-order.","speaker":null,"is_sponsor":0},{"start_s":246.19,"end_s":248.49,"text":"Hold your horses there, Tom.","speaker":null,"is_sponsor":0},{"start_s":248.49,"end_s":249.17,"text":"Yeah, it's awesome.","speaker":null,"is_sponsor":0},{"start_s":249.17,"end_s":252.03,"text":"But the number of links provided on RTX","speaker":null,"is_sponsor":0},{"start_s":252.03,"end_s":253.45,"text":"is relatively minimal,","speaker":null,"is_sponsor":0},{"start_s":253.45,"end_s":258.45,"text":"and the RTX cards only support SLI over the NVLink bus.","speaker":null,"is_sponsor":0},{"start_s":259.61,"end_s":264.01,"text":"So there will be no fancy resource pooling going on here.","speaker":null,"is_sponsor":0},{"start_s":264.01,"end_s":268.75,"text":"So our plan today then is to take our Quadro GP100s","speaker":null,"is_sponsor":0},{"start_s":268.75,"end_s":271.15,"text":"and run them both in compute mode,","speaker":null,"is_sponsor":0},{"start_s":271.15,"end_s":273.45,"text":"which actually disables the graphics engine,","speaker":null,"is_sponsor":0},{"start_s":273.45,"end_s":274.81,"text":"like we couldn't plug a display","speaker":null,"is_sponsor":0},{"start_s":274.81,"end_s":276.73,"text":"into these things right now if we tried,","speaker":null,"is_sponsor":0},{"start_s":276.73,"end_s":278.47,"text":"and in what's called SLI mode,","speaker":null,"is_sponsor":0},{"start_s":278.47,"end_s":281.35,"text":"to look at their gaming performance.","speaker":null,"is_sponsor":0},{"start_s":281.35,"end_s":283.07,"text":"Yes, yes, I know.","speaker":null,"is_sponsor":0},{"start_s":283.07,"end_s":285.75,"text":"This card isn't intended for gaming,","speaker":null,"is_sponsor":0},{"start_s":285.75,"end_s":289.73,"text":"but if you look closely at the spec of it,","speaker":null,"is_sponsor":0},{"start_s":289.73,"end_s":293.61,"text":"it's got HBM2 memory, yes, and more of it,","speaker":null,"is_sponsor":0},{"start_s":293.61,"end_s":298.61,"text":"but it's otherwise actually very similar to the GTX 1080 Ti.","speaker":null,"is_sponsor":0},{"start_s":299.75,"end_s":304.05,"text":"So this is probably as close as we will ever get","speaker":null,"is_sponsor":0},{"start_s":304.05,"end_s":308.05,"text":"to an apples-to-apples comparison between SLI and NVLink.","speaker":null,"is_sponsor":0},{"start_s":308.47,"end_s":313.27,"text":"Since Pascal is likely to be the only generation of products","speaker":null,"is_sponsor":0},{"start_s":313.27,"end_s":316.53,"text":"where both of these technologies are present.","speaker":null,"is_sponsor":0},{"start_s":316.53,"end_s":318.01,"text":"First up, some pre-flight tweaks","speaker":null,"is_sponsor":0},{"start_s":318.01,"end_s":319.55,"text":"to get everything working though.","speaker":null,"is_sponsor":0},{"start_s":319.55,"end_s":322.37,"text":"We needed a Quadro SLI-certified motherboard,","speaker":null,"is_sponsor":0},{"start_s":322.37,"end_s":326.17,"text":"so our ASUS X299 Deluxe with a Core i9-7900X","speaker":null,"is_sponsor":0},{"start_s":326.17,"end_s":327.63,"text":"worked nicely for this.","speaker":null,"is_sponsor":0},{"start_s":327.63,"end_s":331.45,"text":"And to look at NVLink's non-gaming performance,","speaker":null,"is_sponsor":0},{"start_s":331.45,"end_s":333.31,"text":"we needed to configure both cards","speaker":null,"is_sponsor":0},{"start_s":333.31,"end_s":335.75,"text":"in Tesla compute cluster mode,","speaker":null,"is_sponsor":0},{"start_s":335.75,"end_s":338.39,"text":"which we can check by going ahead and running...","speaker":null,"is_sponsor":0},{"start_s":338.47,"end_s":341.37,"text":"this command in the Windows PowerShell.","speaker":null,"is_sponsor":0},{"start_s":341.37,"end_s":342.79,"text":"So you can see right here,","speaker":null,"is_sponsor":0},{"start_s":342.79,"end_s":346.23,"text":"links one to three, or zero to three, excuse me,","speaker":null,"is_sponsor":0},{"start_s":346.23,"end_s":348.09,"text":"or one to, whatever the point is,","speaker":null,"is_sponsor":0},{"start_s":348.09,"end_s":352.61,"text":"they're all running, and that's good.","speaker":null,"is_sponsor":0},{"start_s":352.61,"end_s":355.93,"text":"Unfortunately, many of our benchmarks","speaker":null,"is_sponsor":0},{"start_s":355.93,"end_s":359.65,"text":"actually didn't cooperate very well","speaker":null,"is_sponsor":0},{"start_s":359.65,"end_s":363.74,"text":"with this particular setup,","speaker":null,"is_sponsor":0},{"start_s":363.74,"end_s":367.58,"text":"though the latest experimental Blender build managed it,","speaker":null,"is_sponsor":0},{"start_s":367.58,"end_s":368.42,"text":"and, whew!","speaker":null,"is_sponsor":0},{"start_s":369.36,"end_s":374.11,"text":"The results pretty much speak for themselves.","speaker":null,"is_sponsor":0},{"start_s":374.11,"end_s":377.2,"text":"Three and a half minutes for Gooseberry?","speaker":null,"is_sponsor":0},{"start_s":377.2,"end_s":380.7,"text":"20 seconds for BMW?","speaker":null,"is_sponsor":0},{"start_s":380.7,"end_s":382.3,"text":"In spite of these tests","speaker":null,"is_sponsor":0},{"start_s":382.3,"end_s":384.96,"text":"not being particularly memory intensive,","speaker":null,"is_sponsor":0},{"start_s":384.96,"end_s":388.18,"text":"we are seeing a clear advantage here.","speaker":null,"is_sponsor":0},{"start_s":388.18,"end_s":391.66,"text":"As for Luxmark's lower OpenCL performance scaling,","speaker":null,"is_sponsor":0},{"start_s":391.66,"end_s":395.32,"text":"that suggests that CUDA is a necessary ingredient","speaker":null,"is_sponsor":0},{"start_s":395.32,"end_s":398.26,"text":"if we wanna take full advantage of NVLink.","speaker":null,"is_sponsor":0},{"start_s":398.26,"end_s":399.8,"text":"Big surprise, of course.","speaker":null,"is_sponsor":0},{"start_s":400.1,"end_s":401.48,"text":"Not all there is to it, though.","speaker":null,"is_sponsor":0},{"start_s":401.48,"end_s":403.9,"text":"Remember how NVLink allows us to utilize","speaker":null,"is_sponsor":0},{"start_s":403.9,"end_s":405.96,"text":"all of the available memory on our cards","speaker":null,"is_sponsor":0},{"start_s":405.96,"end_s":408.47,"text":"as though they were one big card?","speaker":null,"is_sponsor":0},{"start_s":408.47,"end_s":410.31,"text":"Well, because of that,","speaker":null,"is_sponsor":0},{"start_s":410.31,"end_s":412.87,"text":"we can now work with much larger data sets","speaker":null,"is_sponsor":0},{"start_s":412.87,"end_s":415.93,"text":"than would have been possible on smaller configurations.","speaker":null,"is_sponsor":0},{"start_s":415.93,"end_s":418.61,"text":"And trust us, we tried on smaller configurations.","speaker":null,"is_sponsor":0},{"start_s":418.61,"end_s":422.61,"text":"You can see here, even our twinned GP100s","speaker":null,"is_sponsor":0},{"start_s":422.61,"end_s":425.13,"text":"couldn't handle this particular workload.","speaker":null,"is_sponsor":0},{"start_s":425.13,"end_s":429.87,"text":"So, it's time to bring out the big guns.","speaker":null,"is_sponsor":0},{"start_s":429.87,"end_s":434.41,"text":"Our GV100s with their new NVLink bridges","speaker":null,"is_sponsor":0},{"start_s":434.41,"end_s":439.41,"text":"will give us a total of 64 gigs of HBM2 memory.","speaker":null,"is_sponsor":0},{"start_s":439.71,"end_s":441.75,"text":"That's more than the system memory","speaker":null,"is_sponsor":0},{"start_s":441.75,"end_s":444.17,"text":"of even many workstations.","speaker":null,"is_sponsor":0},{"start_s":444.17,"end_s":445.81,"text":"And there it is.","speaker":null,"is_sponsor":0},{"start_s":445.81,"end_s":449.63,"text":"Our GV100s handle this just fine.","speaker":null,"is_sponsor":0},{"start_s":449.63,"end_s":453.91,"text":"So, that's super impressive and extremely useful","speaker":null,"is_sponsor":0},{"start_s":453.91,"end_s":457.63,"text":"for people with huge data sets.","speaker":null,"is_sponsor":0},{"start_s":457.63,"end_s":459.41,"text":"But the real thing we were after here was","speaker":null,"is_sponsor":0},{"start_s":459.41,"end_s":464.41,"text":"evaluating the SLI mode that is coming with the RTX series.","speaker":null,"is_sponsor":0},{"start_s":464.57,"end_s":468.17,"text":"So, and here we go.","speaker":null,"is_sponsor":0},{"start_s":468.17,"end_s":470.99,"text":"So, in a massive surprise to no one,","speaker":null,"is_sponsor":0},{"start_s":470.99,"end_s":475.99,"text":"the GV100s are the fastest solution on the block, for now.","speaker":null,"is_sponsor":0},{"start_s":476.41,"end_s":479.59,"text":"In SLI, even at 4K Ultra,","speaker":null,"is_sponsor":0},{"start_s":479.59,"end_s":484.03,"text":"the average frame rate never dipped below 60, which is huge.","speaker":null,"is_sponsor":0},{"start_s":484.03,"end_s":487.51,"text":"Nothing else can even come close to claiming that.","speaker":null,"is_sponsor":0},{"start_s":487.51,"end_s":489.25,"text":"What's more interesting, though,","speaker":null,"is_sponsor":0},{"start_s":489.25,"end_s":492.95,"text":"is when we look at the scaling figures side by side.","speaker":null,"is_sponsor":0},{"start_s":492.95,"end_s":495.91,"text":"So, our GP100s here,","speaker":null,"is_sponsor":0},{"start_s":495.91,"end_s":500.75,"text":"these guys seem to scale better than the GV100s","speaker":null,"is_sponsor":0},{"start_s":500.75,"end_s":502.29,"text":"in gaming and productivity,","speaker":null,"is_sponsor":0},{"start_s":502.29,"end_s":504.59,"text":"giving them the best scaling overall,","speaker":null,"is_sponsor":0},{"start_s":504.59,"end_s":508.21,"text":"which may suggest some kind of CPU bottleneck.","speaker":null,"is_sponsor":0},{"start_s":508.21,"end_s":510.51,"text":"As for the GTX 1080 Ti,","speaker":null,"is_sponsor":0},{"start_s":510.51,"end_s":513.53,"text":"well, there's huge gains to be made in gaming,","speaker":null,"is_sponsor":0},{"start_s":513.53,"end_s":515.85,"text":"but not as much in productivity.","speaker":null,"is_sponsor":0},{"start_s":515.85,"end_s":518.57,"text":"So, as you might expect with anything new,","speaker":null,"is_sponsor":0},{"start_s":519.25,"end_s":522.87,"text":"I think SLI doesn't scale the same way for everything,","speaker":null,"is_sponsor":0},{"start_s":522.87,"end_s":526.49,"text":"but it does look to be a pretty decent improvement","speaker":null,"is_sponsor":0},{"start_s":526.49,"end_s":528.21,"text":"over traditional SLI,","speaker":null,"is_sponsor":0},{"start_s":528.21,"end_s":531.75,"text":"about 10 to 23% better by our measure,","speaker":null,"is_sponsor":0},{"start_s":531.75,"end_s":533.93,"text":"with the potential to dramatically improve","speaker":null,"is_sponsor":0},{"start_s":533.93,"end_s":536.97,"text":"undesirable behavior like micro-stuttering as well,","speaker":null,"is_sponsor":0},{"start_s":536.97,"end_s":539.97,"text":"or even enable more than two-way SLI","speaker":null,"is_sponsor":0},{"start_s":539.97,"end_s":542.49,"text":"with decent scaling in the future.","speaker":null,"is_sponsor":0},{"start_s":542.49,"end_s":545.97,"text":"That is, depending on how much NVIDIA decides to neuter it","speaker":null,"is_sponsor":0},{"start_s":545.97,"end_s":548.65,"text":"compared to its professional-grade cousin.","speaker":null,"is_sponsor":0},{"start_s":548.65,"end_s":549.01,"text":"You, too, never know.","speaker":null,"is_sponsor":0},{"start_s":549.01,"end_s":550.07,"text":"You never know with those guys.","speaker":null,"is_sponsor":0},{"start_s":550.07,"end_s":551.99,"text":"I mean, one thing we discovered","speaker":null,"is_sponsor":0},{"start_s":551.99,"end_s":553.65,"text":"in the course of our testing for this video","speaker":null,"is_sponsor":0},{"start_s":553.65,"end_s":556.41,"text":"is that the new NVLink bridges here","speaker":null,"is_sponsor":0},{"start_s":556.41,"end_s":560.77,"text":"don't work with the old NVLink cards, even the pro ones.","speaker":null,"is_sponsor":0},{"start_s":560.77,"end_s":563.31,"text":"So, NVIDIA told us something about","speaker":null,"is_sponsor":0},{"start_s":563.31,"end_s":566.33,"text":"consumer NVLink bridges having fewer pins,","speaker":null,"is_sponsor":0},{"start_s":566.33,"end_s":569.05,"text":"or more importantly, a slightly different pin-out,","speaker":null,"is_sponsor":0},{"start_s":569.05,"end_s":570.59,"text":"but could they have made it work?","speaker":null,"is_sponsor":0},{"start_s":570.59,"end_s":571.43,"text":"I don't know.","speaker":null,"is_sponsor":0},{"start_s":571.43,"end_s":572.33,"text":"I don't know with those guys.","speaker":null,"is_sponsor":0},{"start_s":572.33,"end_s":576.15,"text":"Either way, NVLink has lots of potential","speaker":null,"is_sponsor":0},{"start_s":576.15,"end_s":578.31,"text":"and looks like a significant hardware upgrade","speaker":null,"is_sponsor":0},{"start_s":578.31,"end_s":579.55,"text":"that should only improve","speaker":null,"is_sponsor":0},{"start_s":579.55,"end_s":582.33,"text":"as the drivers themselves continue to improve.","speaker":null,"is_sponsor":0},{"start_s":582.33,"end_s":586.65,"text":"So maybe, just maybe, SLI isn't dead.","speaker":null,"is_sponsor":0},{"start_s":586.65,"end_s":588.45,"text":"Yet.","speaker":null,"is_sponsor":0},{"start_s":588.45,"end_s":590.25,"text":"Maybe.","speaker":null,"is_sponsor":0},{"start_s":590.25,"end_s":591.31,"text":"But you know what's not maybe?","speaker":null,"is_sponsor":0},{"start_s":591.31,"end_s":594.33,"text":"FreshBooks, being the small business accounting software","speaker":null,"is_sponsor":0},{"start_s":594.33,"end_s":596.51,"text":"custom-built for how you wanna work.","speaker":null,"is_sponsor":0},{"start_s":596.51,"end_s":598.35,"text":"If you're a freelancer or a small business owner,","speaker":null,"is_sponsor":0},{"start_s":598.35,"end_s":600.11,"text":"you need to check out FreshBooks.","speaker":null,"is_sponsor":0},{"start_s":600.11,"end_s":601.85,"text":"It's the simple way to be more productive,","speaker":null,"is_sponsor":0},{"start_s":601.85,"end_s":604.31,"text":"more organized, and to get paid faster.","speaker":null,"is_sponsor":0},{"start_s":604.31,"end_s":606.61,"text":"You can create and send professional-looking invoices","speaker":null,"is_sponsor":0},{"start_s":606.61,"end_s":608.07,"text":"in less than 30 seconds.","speaker":null,"is_sponsor":0},{"start_s":608.07,"end_s":609.41,"text":"You can set up online payments,","speaker":null,"is_sponsor":0},{"start_s":609.41,"end_s":610.57,"text":"with just a couple of clicks,","speaker":null,"is_sponsor":0},{"start_s":610.57,"end_s":612.47,"text":"to get paid up to four days faster,","speaker":null,"is_sponsor":0},{"start_s":612.47,"end_s":614.91,"text":"and you don't have to take my word for it.","speaker":null,"is_sponsor":0},{"start_s":614.91,"end_s":617.29,"text":"Go try FreshBooks for 30 days for free","speaker":null,"is_sponsor":0},{"start_s":617.29,"end_s":619.45,"text":"at freshbooks.com slash techtips.","speaker":null,"is_sponsor":0},{"start_s":619.45,"end_s":620.77,"text":"Just enter Linus Tech Tips","speaker":null,"is_sponsor":0},{"start_s":620.77,"end_s":622.27,"text":"in the How You Heard About Us section,","speaker":null,"is_sponsor":0},{"start_s":622.27,"end_s":624.45,"text":"so they'll know who sent you.","speaker":null,"is_sponsor":0},{"start_s":624.45,"end_s":625.43,"text":"So thanks for watching, guys.","speaker":null,"is_sponsor":0},{"start_s":625.43,"end_s":627.23,"text":"If this video sucked, you know what to do.","speaker":null,"is_sponsor":0},{"start_s":627.23,"end_s":628.91,"text":"But if it was awesome, get subscribed,","speaker":null,"is_sponsor":0},{"start_s":628.91,"end_s":632.33,"text":"hit that like button, or check out the NVLink.","speaker":null,"is_sponsor":0},{"start_s":632.33,"end_s":633.99,"text":"Oh, Lordy, that's awful.","speaker":null,"is_sponsor":0},{"start_s":633.99,"end_s":635.91,"text":"Anthony, come on.","speaker":null,"is_sponsor":0},{"start_s":635.91,"end_s":637.25,"text":"To where to buy the stuff we featured","speaker":null,"is_sponsor":0},{"start_s":637.25,"end_s":638.33,"text":"in the video description.","speaker":null,"is_sponsor":0},{"start_s":638.33,"end_s":639.25,"text":"Also, NVLink.","speaker":null,"is_sponsor":0},{"start_s":639.41,"end_s":640.25,"text":"NVLinked.","speaker":null,"is_sponsor":0},{"start_s":640.25,"end_s":641.45,"text":"In the description is our merch store,","speaker":null,"is_sponsor":0},{"start_s":641.45,"end_s":643.87,"text":"which has cool shirts, and our community forum,","speaker":null,"is_sponsor":0},{"start_s":643.87,"end_s":645.37,"text":"which you should totally join.","speaker":null,"is_sponsor":0},{"start_s":648.54,"end_s":650.3,"text":"She made me do it twice.","speaker":null,"is_sponsor":0},{"start_s":650.3,"end_s":652.42,"text":"I mean, it's really, it's my fault.","speaker":null,"is_sponsor":0},{"start_s":652.42,"end_s":654.42,"text":"I'll read anything on the teleprompter.","speaker":null,"is_sponsor":0}],"full_text":"When NVIDIA released SLI, it enabled gamers to enjoy next generation levels of quality today. Assuming of course, that they could afford literally twice as many graphics cards. But as monitor resolutions have grown, the scalable part of the scalable link interface, which has been with us for over a decade, hasn't been able to keep up in spite of fancy, high bandwidth bridges like this one. Meanwhile, over on the professional side of things, NVIDIA has been pushing a newer inter-GPU communication protocol called NVLink. This essentially turns SLI up to 11. But why would you, the general consumer, care about that? Well, because NVLink is coming to consumers with the GeForce RTX series. So it is time then to ask the big question. Does it make gaming better? Whew, that's a lot of hardware. Speaking of big questions, have you tried GlassWire? Detect malware and block badly behaving apps on your PC or Android device. Use offer code Linus to get 25% off GlassWire 2.0 at the link below. So one of the first things you'll notice about a card equipped with NVLink is just how big the connector fingers are compared to traditional SLI. They are more than three times as wide with way, way more pins. Like seriously, a single NVLink finger is wider than the entire SLI connector setup. It almost even looks like they're little PCI Express connectors, which, as we're about to see, isn't by accident. So the way that SLI works is actually a lot like, oh, here, I have a good prep for this. It's actually a lot like the older SCSI and IDE. One card functions as the master in the relationship and the other one as a slave, or in the case of multiple other cards, they would all then be slaves. So that means that because the master alone is directing the workload for those slave cards with at best two gigabytes per second of bandwidth using one of NVIDIA's high bandwidth bridges, you've got enough for the render results to be returned to the master and honestly, not a whole lot more. This is the reason why you can't simply add together the memory of your SLI graphics cards, taking two 11 gig cards and saying, well, I've got 22 gigs of RAM now. And the same is true for Team Red's Crossfire. By contrast, NVLink is bi-directional and it's configured as a mesh, which means that no one card is the master, and there are no slaves. Think of it more like if you were plugging computers into a router or a switch. So this, along with the extra pins and newer signaling protocol, gives these cards a lot more bandwidth, more than even PCI Express, at a total of up to 160 to 300 gigabytes per second. That kind of speed lets them pool resources, that kind of speed lets them pool resources, that kind of speed lets them pool resources, in a way that allows access to each card's memory and CUDA cores as though they were a single card. And that's perfect for the scientific and high-end render stations that NVIDIA has traditionally targeted with NVLink. Now, you might be thinking to yourself, awesome, NVLink is coming to GeForce RTX cards. We're gonna get those benefits. I'm doubling my pre-order. Hold your horses there, Tom. Yeah, it's awesome. But the number of links provided on RTX is relatively minimal, and the RTX cards only support SLI over the NVLink bus. So there will be no fancy resource pooling going on here. So our plan today then is to take our Quadro GP100s and run them both in compute mode, which actually disables the graphics engine, like we couldn't plug a display into these things right now if we tried, and in what's called SLI mode, to look at their gaming performance. Yes, yes, I know. This card isn't intended for gaming, but if you look closely at the spec of it, it's got HBM2 memory, yes, and more of it, but it's otherwise actually very similar to the GTX 1080 Ti. So this is probably as close as we will ever get to an apples-to-apples comparison between SLI and NVLink. Since Pascal is likely to be the only generation of products where both of these technologies are present. First up, some pre-flight tweaks to get everything working though. We needed a Quadro SLI-certified motherboard, so our ASUS X299 Deluxe with a Core i9-7900X worked nicely for this. And to look at NVLink's non-gaming performance, we needed to configure both cards in Tesla compute cluster mode, which we can check by going ahead and running... this command in the Windows PowerShell. So you can see right here, links one to three, or zero to three, excuse me, or one to, whatever the point is, they're all running, and that's good. Unfortunately, many of our benchmarks actually didn't cooperate very well with this particular setup, though the latest experimental Blender build managed it, and, whew! The results pretty much speak for themselves. Three and a half minutes for Gooseberry? 20 seconds for BMW? In spite of these tests not being particularly memory intensive, we are seeing a clear advantage here. As for Luxmark's lower OpenCL performance scaling, that suggests that CUDA is a necessary ingredient if we wanna take full advantage of NVLink. Big surprise, of course. Not all there is to it, though. Remember how NVLink allows us to utilize all of the available memory on our cards as though they were one big card? Well, because of that, we can now work with much larger data sets than would have been possible on smaller configurations. And trust us, we tried on smaller configurations. You can see here, even our twinned GP100s couldn't handle this particular workload. So, it's time to bring out the big guns. Our GV100s with their new NVLink bridges will give us a total of 64 gigs of HBM2 memory. That's more than the system memory of even many workstations. And there it is. Our GV100s handle this just fine. So, that's super impressive and extremely useful for people with huge data sets. But the real thing we were after here was evaluating the SLI mode that is coming with the RTX series. So, and here we go. So, in a massive surprise to no one, the GV100s are the fastest solution on the block, for now. In SLI, even at 4K Ultra, the average frame rate never dipped below 60, which is huge. Nothing else can even come close to claiming that. What's more interesting, though, is when we look at the scaling figures side by side. So, our GP100s here, these guys seem to scale better than the GV100s in gaming and productivity, giving them the best scaling overall, which may suggest some kind of CPU bottleneck. As for the GTX 1080 Ti, well, there's huge gains to be made in gaming, but not as much in productivity. So, as you might expect with anything new, I think SLI doesn't scale the same way for everything, but it does look to be a pretty decent improvement over traditional SLI, about 10 to 23% better by our measure, with the potential to dramatically improve undesirable behavior like micro-stuttering as well, or even enable more than two-way SLI with decent scaling in the future. That is, depending on how much NVIDIA decides to neuter it compared to its professional-grade cousin. You, too, never know. You never know with those guys. I mean, one thing we discovered in the course of our testing for this video is that the new NVLink bridges here don't work with the old NVLink cards, even the pro ones. So, NVIDIA told us something about consumer NVLink bridges having fewer pins, or more importantly, a slightly different pin-out, but could they have made it work? I don't know. I don't know with those guys. Either way, NVLink has lots of potential and looks like a significant hardware upgrade that should only improve as the drivers themselves continue to improve. So maybe, just maybe, SLI isn't dead. Yet. Maybe. But you know what's not maybe? FreshBooks, being the small business accounting software custom-built for how you wanna work. If you're a freelancer or a small business owner, you need to check out FreshBooks. It's the simple way to be more productive, more organized, and to get paid faster. You can create and send professional-looking invoices in less than 30 seconds. You can set up online payments, with just a couple of clicks, to get paid up to four days faster, and you don't have to take my word for it. Go try FreshBooks for 30 days for free at freshbooks.com slash techtips. Just enter Linus Tech Tips in the How You Heard About Us section, so they'll know who sent you. So thanks for watching, guys. If this video sucked, you know what to do. But if it was awesome, get subscribed, hit that like button, or check out the NVLink. Oh, Lordy, that's awful. Anthony, come on. To where to buy the stuff we featured in the video description. Also, NVLink. NVLinked. In the description is our merch store, which has cool shirts, and our community forum, which you should totally join. She made me do it twice. I mean, it's really, it's my fault. I'll read anything on the teleprompter."}