WEBVTT 00:00:00.000 --> 00:00:09.660 what gets FPS gpus what gets more FPS 00:00:06.299 --> 00:00:11.340 more gpus obviously everybody thinks so 00:00:09.660 --> 00:00:16.619 and manufacturers have been working on this technology since the late 90s so 00:00:13.920 --> 00:00:19.859 then what the heck happened to it I'm gonna let you know a little secret 00:00:17.760 --> 00:00:23.279 come here hide 00:00:21.720 --> 00:00:27.060 future now that might seem like a bold 00:00:25.140 --> 00:00:31.740 prediction given that NVIDIA just completely removes the multi-GPU links 00:00:29.580 --> 00:00:36.480 from their latest generation even on the premium workstation cards but there is a 00:00:34.320 --> 00:00:42.239 clear course set for better performance through multi-GPU Graphics Solutions and 00:00:40.020 --> 00:00:46.680 the most bizarre part is that it's Apple who ultimately did it to properly see 00:00:44.820 --> 00:00:53.219 where we're going sometimes we have to visit the past which is why we got our 00:00:48.960 --> 00:00:57.180 hands on this legendary piece of Kit the 00:00:53.219 --> 00:01:00.539 quantum 3D Mercury brick is an 8 GPU 00:00:57.180 --> 00:01:02.460 design from 25 years ago that has a lot 00:01:00.539 --> 00:01:07.200 more in common with your future gaming rig than you probably realize like you 00:01:05.460 --> 00:01:12.060 realize I'm about to tell you about our sponsor zoho1 unify all your business 00:01:10.080 --> 00:01:16.200 processes into one comprehensive platform and improve productivity with 00:01:14.159 --> 00:01:20.460 Zoho one you can lay all concerns about integration and compatibility to rest 00:01:18.420 --> 00:01:24.740 follow the link below for a free 30-day trial Our Story begins today with the 00:01:22.619 --> 00:01:28.080 quantum 3D obsidian 100sb-440v that Ross tregamba 00:01:26.880 --> 00:01:33.060 thoughtfully brought along when we shot our Voodoo 5 6000 video the name might 00:01:30.900 --> 00:01:38.159 not sound like much but it's a great example of something that the original 00:01:34.740 --> 00:01:41.880 3dfx Voodoo one was capable of even if 00:01:38.159 --> 00:01:44.579 it rarely did it dual GPU operation in 00:01:41.880 --> 00:01:49.079 this case on a single board now these ones were intended for desktop PCS but 00:01:47.100 --> 00:01:54.899 the primary purpose of cards like this was to drive arcade machines more on 00:01:52.259 --> 00:02:00.479 that later because first I want to talk about the brick this extremely rare 00:01:58.140 --> 00:02:07.320 piece of computing history uses a four card setup with each card equipped with 00:02:03.360 --> 00:02:08.880 two Voodoo 2 gpus that is eight gpus 00:02:07.320 --> 00:02:13.980 running all together on the same workload now I know what you're thinking 00:02:11.720 --> 00:02:19.020 that looks like it would have a tough time fitting in a PC that also needs 00:02:16.319 --> 00:02:25.140 dedicated cards for sound networking and yes 2D Graphics to run the desktop 00:02:22.020 --> 00:02:27.480 because the voodoo 2 was 3D only and 00:02:25.140 --> 00:02:31.739 you'd be right about that but again more on that later for now 00:02:29.220 --> 00:02:35.099 it's time to see this thing in action I was going to be the one to install the 00:02:33.360 --> 00:02:40.020 brick in this simulator chassis from 1999 but understandably the owner didn't 00:02:37.560 --> 00:02:44.879 trust me to do it without dropping it so um I didn't but we've got some b-roll of 00:02:42.180 --> 00:02:49.140 how it goes in it's wild I mean all four cards need to go into their own PCI 00:02:46.739 --> 00:02:54.000 slots then it's got to be wired up to a VGA out and in then it's got to be wired 00:02:52.080 --> 00:02:57.480 up to a VGA pass-through at the back of the chassis then these two cables are 00:02:56.099 --> 00:03:03.360 wired up to the back for a feature called swap lock that allows more than 00:03:00.239 --> 00:03:07.019 one of these than twenty thousand dollar 00:03:03.360 --> 00:03:08.940 machines to be utilized at once the rest 00:03:07.019 --> 00:03:12.959 of these specs are high end even if relatively mundane we've got dual 00:03:10.800 --> 00:03:18.900 Pentium 3500 megahertz processors what appears to be an Intel first party 00:03:15.300 --> 00:03:21.060 extended motherboard a mere 425 watt 00:03:18.900 --> 00:03:25.860 power supply from the legendary PC power and cooling we've got 256 Megs of RAM a 00:03:24.000 --> 00:03:31.019 gigantic sound card that's as big as some gpus today and then a ton of 00:03:28.739 --> 00:03:35.760 cooling is especially for the time dual 92 millimeter fans and then a couple 00:03:33.420 --> 00:03:40.560 more 80 mils that blow directly down onto the heatsinks over the gpus with 00:03:37.920 --> 00:03:43.459 this what looks very handmade shroud let's fire it up 00:03:43.680 --> 00:03:46.680 oh 00:03:46.799 --> 00:03:53.220 this is part of the simulator experience it needs to sound like it's taking off 00:03:51.299 --> 00:03:57.000 wait there's no event for those 80 millimeter fans 00:03:55.379 --> 00:04:02.819 they're just yeah it's got enough spacing in there for just okay all right 00:04:00.239 --> 00:04:06.659 Windows NT let's go we have a few options for things we can run we could 00:04:04.620 --> 00:04:09.959 play games on it which obviously is the first thing I would have done if I 00:04:07.920 --> 00:04:12.959 worked at a military installation in one of these showtimes how fast does it run 00:04:11.700 --> 00:04:16.739 halfway but we actually also have some simulator 00:04:15.120 --> 00:04:22.139 software cool it's so weird seeing a Sierra 00:04:18.780 --> 00:04:24.540 splash screen when you fire up half-life 00:04:22.139 --> 00:04:33.800 you mean valve right right about okay all the geometry what the what that 00:04:29.580 --> 00:04:36.419 stop stop buddy check out the pixelation 00:04:33.800 --> 00:04:41.340 whoa the first thing I noticed was that compared to the voodoo 5 6000 the frame 00:04:38.820 --> 00:04:47.460 rate is obviously lower what I didn't think to look at 00:04:42.900 --> 00:04:52.080 is house move it is what the 00:04:47.460 --> 00:04:54.360 this was possible in 1999 00:04:52.080 --> 00:04:59.400 I think that's functionally 16 times anti-aliasing this looks 00:04:57.300 --> 00:05:03.000 modern cleaner than modern 00:05:01.259 --> 00:05:09.000 it looks like a much higher resolution than it is that is wild 00:05:06.479 --> 00:05:13.620 like I'm specifically going out of my way to put these straight lines at just 00:05:11.820 --> 00:05:18.060 you know ever so slight an angle so that I can see the stairs 00:05:15.300 --> 00:05:22.380 it's just not there it really makes the crappy old texture stand out doesn't it 00:05:19.979 --> 00:05:27.479 oh yeah like if this had the horsepower to run modern games with modern 00:05:24.419 --> 00:05:29.720 resolution textures it would look so 00:05:27.479 --> 00:05:29.720 good 00:05:30.500 --> 00:05:33.500 blah blah 00:05:34.139 --> 00:05:40.620 blah blah got him it's Hardware level anti-aliasing okay that's pretty wild I 00:05:39.180 --> 00:05:44.759 think I do kind of need to see it without the hardware anti-aliasing then 00:05:42.960 --> 00:05:49.800 okay so how do you uh is that something you can change while the game is running 00:05:46.320 --> 00:05:52.740 yes oh no way yeah so no wait wait what 00:05:49.800 --> 00:05:57.600 so this is a Medusa cable uh-huh so it's got a 26 spin guy here two bgas so what 00:05:56.100 --> 00:06:03.240 we're gonna do is we're gonna bypass this by just plugging into one of the 00:06:00.300 --> 00:06:08.940 boards individually so in this case each board is a dual GPU GPU and the other 00:06:06.840 --> 00:06:14.100 three boards are just contributing to more anti-aliasing correct yeah so the 00:06:12.120 --> 00:06:18.600 bridge is basically combining the image it's basically four instances of voodoo 00:06:16.440 --> 00:06:25.440 to SLI okay basically what it is so I'll get the same FPS correct but it'll look 00:06:22.020 --> 00:06:27.360 like crap exactly 00:06:25.440 --> 00:06:34.740 we're both going pretty much the same place slightly different stages I can confirm 00:06:31.800 --> 00:06:39.720 anti-aliasing no longer working now it still looks smoother than it 00:06:36.960 --> 00:06:44.460 would on an LCD just because CRTs have they're not really pixels they're more 00:06:42.000 --> 00:06:48.120 like dots so they're they're not squared at the edges 00:06:45.600 --> 00:06:52.860 but you can really see the difference here oh look especially on objects that 00:06:51.060 --> 00:06:57.960 are farther away where the level of detail is lower like those picture 00:06:54.300 --> 00:06:59.759 frames yeah that's a big yikes 00:06:57.960 --> 00:07:05.160 and the shimmering as I'm moving back and forth that is a big difference not 00:07:03.780 --> 00:07:09.900 the kind of difference the average home user would pay for but a difference 00:07:07.440 --> 00:07:14.180 nonetheless let's try some simulation stuff it's a non-interactive right I 00:07:12.419 --> 00:07:18.800 guess I wouldn't have whatever control cockpit they would use for something 00:07:16.259 --> 00:07:18.800 like this 00:07:20.039 --> 00:07:23.539 and that is smooth 00:07:23.639 --> 00:07:30.300 60fps in 1999 00:07:27.840 --> 00:07:35.759 that's the thing we got to remember hold on I'm not I'm not done I'm not done 00:07:32.639 --> 00:07:38.900 checking out how smooth the terrain 00:07:35.759 --> 00:07:41.520 looks like yeah we're seeing that 00:07:38.900 --> 00:07:47.039 whether that's like a frame pacing issue or whether that's just a vsync issue 00:07:45.240 --> 00:07:52.500 isn't it pretty incredible how well the AAA Works in this thing 00:07:48.740 --> 00:07:54.960 anti-aliasing is unreal 00:07:52.500 --> 00:07:58.440 like look at these particle effects like yeah they're not amazing or 00:07:56.940 --> 00:08:02.639 whatever but they exist 00:08:01.020 --> 00:08:08.220 there's just no other way to put it this machine is mind-blowing but what's most 00:08:05.759 --> 00:08:13.979 mind-blowing about it at least to me is that the maker of the brick Quantum 3D 00:08:10.740 --> 00:08:15.900 still exists as a producer of simulation 00:08:13.979 --> 00:08:19.620 hardware and software if you remember the Evans and Sutherland Sim Fusion card 00:08:18.240 --> 00:08:22.860 that we looked at a couple of years ago these guys probably would have been in 00:08:21.479 --> 00:08:29.580 competition with that at some point going back to the days of the brick though they sometimes sold enhanced 00:08:26.759 --> 00:08:34.979 versions of 3dfx's single GPU cards to enthusiasts with extras like additional 00:08:32.520 --> 00:08:39.000 video memory and texture units but their main business was to act as 3dfx's 00:08:37.020 --> 00:08:43.200 arcade and simulation division until they were spun off as their own company 00:08:40.500 --> 00:08:47.339 in 1997. with that history it makes a lot of sense that they had the know-how 00:08:44.700 --> 00:08:52.140 to create Exotic Solutions like these out of voodoo gpus the first known use 00:08:49.980 --> 00:08:57.240 of a Quantum 3D card was Ice Home Run Derby in 1996 and a lot of high-profile 00:08:55.320 --> 00:09:01.800 games like San Francisco Rush would go on to use them too these were not cheap 00:08:59.640 --> 00:09:06.720 solutions ending up insist is costing anywhere from over a thousand to 00:09:04.019 --> 00:09:10.440 thousands to in excess of ten thousand dollars but while it would have been 00:09:08.580 --> 00:09:14.580 Unthinkable for a home user to spend more than a few hundred dollars on a 00:09:12.360 --> 00:09:19.620 gaming device at the time at the high end Quantum 3DS devices were used for 00:09:16.740 --> 00:09:23.580 military simulation and even down at the low end they could be expected to 00:09:21.060 --> 00:09:26.940 deliver tens of thousands of dollars of revenue for their owners quarter by 00:09:25.200 --> 00:09:31.560 quarter over the lifetime of the machine and it's that lifetime that was one of 00:09:29.580 --> 00:09:36.899 their big advantages because before Quantum 3D it was quite common for 00:09:33.839 --> 00:09:40.080 arcade games to use bespoke Hardware or 00:09:36.899 --> 00:09:42.720 at best a platform like the Neo Geo by 00:09:40.080 --> 00:09:47.220 comparison Quantum 3DS platform was more or less a PC in all but form factor 00:09:45.060 --> 00:09:51.420 which made changing over to a new game even one that requires a hardware 00:09:48.959 --> 00:09:55.920 upgrade much simpler and more cost effective than it had been in the past 00:09:52.860 --> 00:09:58.740 often game data was even stored on a bog 00:09:55.920 --> 00:10:03.300 standard IDE hard drive and because the Glide API that the gpus ran on was 00:10:00.899 --> 00:10:09.240 readily available and had a similar syntax to opengl games were much easier 00:10:06.300 --> 00:10:13.260 to prototype and develop then as now the hardware is only part of the battle and 00:10:11.339 --> 00:10:17.160 the real work is in making it easier for developers to get the most out of it and 00:10:15.180 --> 00:10:21.540 get the most out of it they did the additional power on tap from Quantum 3D 00:10:19.500 --> 00:10:27.660 cards could be used in one of two ways either as a traditional SLI setup where 00:10:24.540 --> 00:10:30.120 the additional GPU or gpus could be used 00:10:27.660 --> 00:10:34.140 to push resolutions or frame rates higher than they would be able to 00:10:31.440 --> 00:10:39.740 otherwise or as was common in the simulator space to have multiple cards 00:10:36.600 --> 00:10:42.240 generate the same image but with rotated 00:10:39.740 --> 00:10:46.440 anti-aliasing patterns this was a way for them to Output ridiculously smooth 00:10:44.220 --> 00:10:50.399 Graphics in the absence of resolution that was high enough to prevent aliasing 00:10:48.300 --> 00:10:55.260 that is how the Mercury brick would have been used and this was actually very 00:10:52.860 --> 00:11:00.240 important because in a simulator hard pixel edges can be very distracting 00:10:57.360 --> 00:11:06.420 unlike our new circuit desk pad on LTT Store stealthy AF now of course with the 00:11:03.480 --> 00:11:10.740 demise of 3dfx came the demise of SLI as the world knew it then but that didn't 00:11:08.459 --> 00:11:14.399 mean that companies like Quantum 3D and Evans in Sutherland were left without 00:11:12.300 --> 00:11:18.839 Hardware to sell multi-GPU development was seen as essential in the Sim Space 00:11:16.680 --> 00:11:23.160 and continued in Earnest even while there wasn't yet an official way to make 00:11:20.820 --> 00:11:28.140 it work for major players be they NVIDIA or ATI in fact Alienware even had a 00:11:26.459 --> 00:11:32.940 project in the works to run multiple gpus before New Age SLI and Crossfire 00:11:31.260 --> 00:11:37.260 came along when those Technologies finally hit the scene though multi-GPU 00:11:35.040 --> 00:11:40.800 came back to the masses in a big way with some very interesting 00:11:38.579 --> 00:11:45.839 implementations early Crossfire for example required a master GPU that 00:11:43.980 --> 00:11:50.220 stitched together the output from multiple gpus rendering different parts 00:11:48.120 --> 00:11:54.180 of the scene with an external dongle this method is called split frame 00:11:52.140 --> 00:11:58.500 rendering and is similar in concept to how old school SLI used to work it 00:11:56.700 --> 00:12:03.000 wasn't ideal though even aside from the hardware complexity this setup limited 00:12:00.540 --> 00:12:06.660 the maximum resolution there was also software Crossfire at the time which 00:12:04.800 --> 00:12:11.579 allowed dongle free operation directly over the PCI Express bus but PCIe wasn't 00:12:09.899 --> 00:12:16.740 as fast then as it is now and it was only really suitable for low end gpus 00:12:13.680 --> 00:12:18.839 newage SLI and Crossfire X meanwhile 00:12:16.740 --> 00:12:23.579 used Bridges between the Cards themselves which is actually quite 00:12:20.940 --> 00:12:27.959 similar to 3dfx's SLI the difference though was that rather than stitching 00:12:25.620 --> 00:12:32.160 together scan lines or quadrants of the screen the cards could communicate with 00:12:29.760 --> 00:12:37.440 each other over the bridge and divide up their work by each rendering alternating 00:12:34.380 --> 00:12:40.920 frames leading to you guessed it a 00:12:37.440 --> 00:12:44.579 return to multi-GPU cards like the GTX 00:12:40.920 --> 00:12:46.680 590 and Radeon 6990 even three-way and 00:12:44.579 --> 00:12:51.360 4-way configurations were possible using multiple Bridges these setups were 00:12:48.839 --> 00:12:55.440 gargantuan and power hungry but they represented the best performance you 00:12:52.980 --> 00:12:59.459 could get at the time and unlike in the past they were available to the general 00:12:57.180 --> 00:13:03.660 consumer now while it's obviously simpler to implement the alternate frame 00:13:01.800 --> 00:13:07.740 rendering approach unfortunately led to the micro stuttering that we became 00:13:05.279 --> 00:13:12.480 familiar with in the 2010s since each GPU might complete its frame in a 00:13:10.200 --> 00:13:16.260 slightly different amount of time the pacing of frame delivery could be all 00:13:14.399 --> 00:13:20.940 over the place resulting in impressive benchmarks but a really poor gaming 00:13:18.660 --> 00:13:24.839 experience some attempts were made to smooth out frame delivery NVIDIA for 00:13:23.100 --> 00:13:28.200 example released a high bandwidth version of the SLI bridge that allowed 00:13:26.579 --> 00:13:32.820 the cards to communicate more quickly but it just wasn't able to keep up with 00:13:30.660 --> 00:13:36.899 advancements in monitor technology both in terms of resolution and refresh rate 00:13:35.279 --> 00:13:42.600 these Niche projects were taking considerable resources from both AMD and 00:13:39.779 --> 00:13:46.139 NVIDIA that along with tepid enthusiasm from game developers who really prefer 00:13:44.399 --> 00:13:51.000 to Target the lowest common denominator ultimately led to the demise of consumer 00:13:48.839 --> 00:13:55.139 multi-GPU it's not that NVIDIA gave up without a fight though they actually 00:13:52.620 --> 00:13:59.100 quietly made an attempt to resurrect SLI in the RTX 2000 generation when they 00:13:57.420 --> 00:14:03.120 explored checkerboard frame rendering unlike split frame rendering we which 00:14:01.079 --> 00:14:07.920 would have both gpus contributing to each displayed frame but would often 00:14:05.279 --> 00:14:12.839 have an ugly scene down the middle CFR would divide the frame into small chunks 00:14:10.260 --> 00:14:17.339 for each card to render then composite the result into a single image in theory 00:14:15.720 --> 00:14:21.600 it could have been the best of all worlds and even avoid NVIDIA having to 00:14:19.680 --> 00:14:27.120 specifically optimize a given game for multi-GPU but they ultimately abandoned 00:14:24.899 --> 00:14:31.019 it for a few reasons first the processing overhead was significant 00:14:28.860 --> 00:14:36.480 which could increase latency and second DirectX 12's explicit multi-GPU model 00:14:34.139 --> 00:14:41.160 showed up around that time which handed multi-GPU control over to the game 00:14:38.519 --> 00:14:45.899 developer entirely offloading it from the driver which game developers 00:14:43.079 --> 00:14:51.120 obviously haven't done anything with but that doesn't mean the multi-GPU dream is 00:14:48.120 --> 00:14:53.639 dead and of all people it's Apple 00:14:51.120 --> 00:14:59.699 showing us the way you see the M1 series of socs has far more in common with our 00:14:56.940 --> 00:15:03.000 Quantum 3D brick than meets the eye have a look at these die shots and you'll see 00:15:01.139 --> 00:15:08.639 what I mean we could think of each of these GPU core blocks on the M1 Pro as 00:15:05.880 --> 00:15:13.139 similar to the GPU chips on our brick with a fast enough interconnect they can 00:15:11.040 --> 00:15:18.420 act as one to render a single output like we see in the M1 Max but then okay 00:15:16.680 --> 00:15:23.279 nothing that we're looking at here is really multi-GPU it's just a bigger 00:15:21.660 --> 00:15:27.959 single die GPU it's only when we move to the M1 Ultra 00:15:25.560 --> 00:15:35.100 that things get really interesting you see the 64 core GPU in the M1 Ultra acts 00:15:32.399 --> 00:15:40.320 for all intents and purposes like a single die solution but Apple disclosed 00:15:38.040 --> 00:15:44.220 in their launch event that is actually a dual die design that uses a 00:15:42.540 --> 00:15:52.260 game-changing packaging solution that they call Ultra Fusion so it's two gpus 00:15:47.420 --> 00:15:54.420 acting as one it's SLI but with 00:15:52.260 --> 00:15:59.399 absolutely none of the downsides how on Earth did they do it well you see the 00:15:56.760 --> 00:16:04.860 problem has always been the interconnect speed see traditionally once you go off 00:16:02.519 --> 00:16:09.120 a single die your bandwidth drops dramatically that's why NVIDIA couldn't 00:16:06.899 --> 00:16:14.699 achieve vram coherency across multiple cards via their SLI Bridges the links 00:16:11.760 --> 00:16:18.839 were just too slow it's also why tricks like split frame or alternate frame 00:16:16.980 --> 00:16:23.760 rendering have always been necessary they couldn't just pull together the GPU 00:16:21.839 --> 00:16:28.440 resources because the left hand would have no idea what the right hand was 00:16:25.260 --> 00:16:31.560 doing in real time but Apple did it and 00:16:28.440 --> 00:16:33.899 both AMD and NVIDIA have also talked 00:16:31.560 --> 00:16:39.120 publicly about the possibility of future GPU designs that use multiple chiplets 00:16:36.839 --> 00:16:44.339 with high-speed interconnects in order to economically scale performance it's 00:16:41.820 --> 00:16:50.940 coming back baby but how does this factor in well the thing is back in 1998 00:16:47.759 --> 00:16:53.940 the bandwidth requirements for inter-GPU 00:16:50.940 --> 00:16:56.279 communication weren't as high so your 00:16:53.940 --> 00:17:00.240 Ultra Fusion interface could look a lot more like this 00:16:58.560 --> 00:17:03.660 it's really funny isn't it how the more things change the more they stay the 00:17:01.860 --> 00:17:08.760 same just like I always do the same segue to our sponsor vessi it's 00:17:06.839 --> 00:17:12.000 springtime which means rain can still show up and ruin your day like an 00:17:10.439 --> 00:17:16.100 annoying neighbor but that's fine because hello we have Bessie shoes 00:17:13.860 --> 00:17:21.360 Bessie claims their shoes are 100 waterproof so we can go step into as 00:17:18.600 --> 00:17:25.260 many puddles as we want they're dymatex technology keeps their shoes light and 00:17:23.280 --> 00:17:29.220 breathable while still being stretchy and comfortable also all of their 00:17:27.240 --> 00:17:33.059 products are vegan and cruelty free maybe you're planning a trip to get away 00:17:31.260 --> 00:17:37.500 from some of this miserable weather don't forget to pack a pair of essies 00:17:35.160 --> 00:17:42.299 their stretchiness means they can easily pack into your luggage or carry-on rain 00:17:39.960 --> 00:17:48.240 or shine here or there it's always a great occasion to wear a pair of vessies 00:17:43.980 --> 00:17:50.400 mon frere go to vesi.com Linus Tech tips 00:17:48.240 --> 00:17:54.419 and get 15 off your purchase with code Linus Tech tips if you guys enjoyed this 00:17:52.919 --> 00:18:00.179 video go check out our coverage of the Evans and Sutherlands Sim Fusion 6500 00:17:57.120 --> 00:18:04.100 Cube for more multi-GPU Madness that 00:18:00.179 --> 00:18:04.100 thing is really cool