WEBVTT 00:00:00.240 --> 00:00:06.400 Oo, really fast card. Hey. Ow. That 00:00:04.319 --> 00:00:12.400 price though. Turning back the clock on NVIDIA's biggest GPU releases, we find 00:00:08.720 --> 00:00:14.799 plenty of wins, but also some L's so big 00:00:12.400 --> 00:00:18.320 that NVIDIA's own marketing team openly mocked them, which got us thinking. Tear 00:00:16.880 --> 00:00:22.320 lists are easy views, right? Sorry, excuse me. Which got us thinking, with 00:00:20.240 --> 00:00:28.000 the benefit of hindsight, what were NVIDIA's best and worst generations? 00:00:25.760 --> 00:00:32.559 Have you been a savvy shopper or did you throw your money at the wrong horses? If 00:00:30.320 --> 00:00:39.760 our only consideration was performance, this would be a pretty simple exercise. 00:00:34.800 --> 00:00:41.840 Best, worst, rest. But there's a lot 00:00:39.760 --> 00:00:46.320 more to delighting gamers than frames per second. So, as we make our way 00:00:43.680 --> 00:00:49.760 through 17 generations of NVIDIA hardware, we're going to be looking at 00:00:47.920 --> 00:00:55.120 the bigger picture with particular attention to efficiency, new features, 00:00:52.640 --> 00:01:00.399 notable controversies, and of course, price. Let me set the stage for our 00:00:57.600 --> 00:01:04.640 first generation. It was 2003, cargo pants for everywhere. I cracked open a 00:01:02.800 --> 00:01:10.479 copy of the computer paper and was shocked to discover that anyone would be 00:01:06.799 --> 00:01:12.400 deranged enough to spend $500 00:01:10.479 --> 00:01:18.320 on a card that does nothing but make games run better. Oh, how naive I was. 00:01:15.920 --> 00:01:23.280 The good news is I kind of dodged a bullet there. To say that NVIDIA's first 00:01:20.799 --> 00:01:27.439 DirectX9 GPUs were controversial would be a gross understatement. Depending on 00:01:25.600 --> 00:01:32.720 the gamer benchmark, performance gains compared to the previous generation were 00:01:29.200 --> 00:01:35.840 up to 55% but could be as low as just 00:01:32.720 --> 00:01:39.280 10%. And that uplift came at the cost of 00:01:35.840 --> 00:01:41.360 noise. And the pain kept coming. NVIDIA 00:01:39.280 --> 00:01:46.399 marketed the FX series cinematic rendering capabilities as CineX. And 00:01:44.400 --> 00:01:51.360 image quality was indeed a big discussion point for this generation, 00:01:48.479 --> 00:01:55.680 just for all the wrong reasons. Not anticipating that new games would use 00:01:53.360 --> 00:02:02.079 the baseline 24-bit color precision mode of DirectX9 rather than 32-bit, FX cards 00:01:59.520 --> 00:02:06.719 ended up falling back to the 16bit mode to increase frame rates. The result was 00:02:04.799 --> 00:02:11.920 a noticeable degradation of image quality. Worse still, NVIDIA attempted 00:02:09.679 --> 00:02:16.879 to slip this little trick into driver version 43.45 00:02:14.000 --> 00:02:21.360 without acknowledging it until the obviously lower image quality forced 00:02:18.800 --> 00:02:26.160 them to come clean. Additionally, the implementation of MSAA on these cards 00:02:23.599 --> 00:02:30.239 was distinctly inferior to ATI's, leading to more pronounced stairstepping 00:02:28.000 --> 00:02:35.760 on the edges of angle geometry. To top it off, ATI's 9700 Pro not only 00:02:33.519 --> 00:02:41.599 performed better, but came out earlier and was the same price as the FX5800 00:02:38.879 --> 00:02:47.120 Ultra. The FX line of GPUs ended up being about as damaging to NVIDIA's 00:02:43.440 --> 00:02:50.560 brand as later FX CPUs were to AMD's. 00:02:47.120 --> 00:02:52.640 Clearly a cursed naming scheme. D tier. 00:02:50.560 --> 00:03:01.040 The good news is that after a less embarrassing mid-generation 5900 series 00:02:56.080 --> 00:03:03.200 refresh of the FX lineup in 2004, Ctier, 00:03:01.040 --> 00:03:08.239 NVIDIA would follow up with a real banger. The performance gains of the 00:03:05.200 --> 00:03:11.519 GeForce 6 generation were outstanding. 00:03:08.239 --> 00:03:13.760 The mid-range 6600 GT outperformed the 00:03:11.519 --> 00:03:18.080 previous flagship for less than half the price. And if you went from flagship to 00:03:16.480 --> 00:03:24.319 flagship, you could double your performance or more. That's right. It 00:03:21.200 --> 00:03:26.959 was in 2004 that SLI made its triumphant 00:03:24.319 --> 00:03:32.080 return from the dead. Early adopters had gotten a taste of this dual GPU wizardry 00:03:29.040 --> 00:03:33.920 under the 3DFX name back in 1998, but it 00:03:32.080 --> 00:03:38.560 would take NVIDIA 4 years after acquiring their rivals to reintroduce 00:03:36.159 --> 00:03:42.879 this technology to a much larger gaming audience. The competition with ATI 00:03:40.879 --> 00:03:47.120 though was tight in this generation. Flagship cards were neck andneck in both 00:03:45.040 --> 00:03:51.760 price and performance with ATI's Crossfire offering similar dual GPU 00:03:49.440 --> 00:03:56.560 capabilities, though Team Red's solution was flakier and required a special 00:03:54.319 --> 00:04:00.879 master card along with a silly external dongle. Furthermore, the best board 00:03:59.040 --> 00:04:05.519 partners were definitely available on the NVIDIA side with many offering 00:04:03.040 --> 00:04:11.439 lifetime warranties. This has got to be S tier. Moving on, as we saw with FX, 00:04:09.040 --> 00:04:18.000 the GeForce 7 family got multiple flagship GPU launches. First came the 00:04:14.400 --> 00:04:19.440 7800 GTX built on the same 110 nanometer 00:04:18.000 --> 00:04:24.240 manufacturing process as its predecessor, but with a larger die, 00:04:21.840 --> 00:04:30.000 higher clock speeds, and graphics optimized GDDR memory, contributing to a 00:04:27.040 --> 00:04:35.919 healthy 60% uplift in performance, even if it did come at a higher price. But 00:04:32.479 --> 00:04:38.000 then we got the 512meg version with 00:04:35.919 --> 00:04:44.479 vastly higher clock speeds and performance and a super misleading name. 00:04:40.960 --> 00:04:46.400 Then we got the 7900 GTX that cranked 00:04:44.479 --> 00:04:51.520 clock speeds to 11 thanks to a more advanced 90 nmter manufacturing process, 00:04:49.040 --> 00:04:58.880 but then barely managed to outperform the 7800 GTX. No, no, not that 7800. 00:04:56.000 --> 00:05:04.320 That one. Oh, and then the 7900 was at a lower price. Do we have an M tier for 00:05:01.199 --> 00:05:06.240 mixed bag? Um, you know what though? 00:05:04.320 --> 00:05:10.160 NVIDIA clearly won every mid-generational battle with ATI this 00:05:08.240 --> 00:05:15.440 time around. This was the last generation to be available for both PCIe 00:05:12.800 --> 00:05:20.080 and AGP slot motherboards, and it was full of fun oddities like the uh sample 00:05:18.240 --> 00:05:27.600 that we have from this generation, the 7950 GX2 that Frankensteined two SLI 00:05:24.320 --> 00:05:31.360 GPUs together into a single PCIe slot. 00:05:27.600 --> 00:05:33.360 Oh yeah, and the 7900 GTO. 00:05:31.360 --> 00:05:37.680 That's right. NVIDIA used to bless us with limited run cards as a way of 00:05:35.440 --> 00:05:41.759 clearing out old or not quite peak performance inventory. And in this case, 00:05:39.840 --> 00:05:47.120 many GTO cards could actually be unlocked to full GTX specifications. It 00:05:44.880 --> 00:05:52.880 was also the last time we saw a flagship GPU offered in a single slot form 00:05:49.120 --> 00:05:54.320 factor. And fun fact, the 7800 GTX was 00:05:52.880 --> 00:06:00.160 used as the starting point for the PlayStation 3's RSX GPU. I'm going with 00:05:57.759 --> 00:06:05.840 A tier this time around. And as for you 7900, 00:06:01.840 --> 00:06:08.400 you get a B. GeForce 8 was, in a word, 00:06:05.840 --> 00:06:13.919 revolutionary. S tier. I'm doing it right now. But we got to talk about it a 00:06:11.440 --> 00:06:22.319 bit more before we move on. First up, a single 8800 GTX was able to outperform a 00:06:17.440 --> 00:06:24.479 pair of 7900 GTXs running an SLI. It is 00:06:22.319 --> 00:06:30.080 still the single biggest generational leap in performance we have ever seen. 00:06:26.880 --> 00:06:34.319 And it cost only $100 more than a single 00:06:30.080 --> 00:06:37.360 7900 GTX, leaving the ATI 2900 XT, 00:06:34.319 --> 00:06:39.919 absolutely in the dust, crying to its 00:06:37.360 --> 00:06:43.280 mama. Its new mama, AMD, who was probably feeling like entering the ring 00:06:41.440 --> 00:06:48.560 with NVIDIA, might not have been a very good idea. The Tesla architectures big 00:06:46.160 --> 00:06:52.800 innovation was doing away with the fixed pixel pipelines of previous micro 00:06:50.240 --> 00:06:57.440 architectures. So instead of separate vertex and pixel shaders, we got unified 00:06:55.520 --> 00:07:02.080 shaders, meaning that any one of them was able to handle any type of shading 00:06:59.440 --> 00:07:06.639 task. This flexibility was extremely important as GPUs made the transition 00:07:04.240 --> 00:07:10.880 from graphics accelerators to the generalpurpose accelerators that power 00:07:08.560 --> 00:07:15.039 workloads like simulation and machine learning today. Also key to that 00:07:13.120 --> 00:07:19.280 transition was the release of the CUDA API, which is still in use today and 00:07:17.199 --> 00:07:24.240 enables developers to run generalpurpose code on GPUs. Less in use today, but 00:07:22.960 --> 00:07:28.800 certainly interesting, was the introduction of three-way SLI for 00:07:26.319 --> 00:07:33.039 enthusiasts with more dollars than cents and plenty of patience for tinkering. 00:07:31.120 --> 00:07:40.000 And if all that wasn't enough, this generation also welcomed NVIDIA's first 00:07:35.280 --> 00:07:42.400 GPU using TSMC's 65nm process. The 8800 00:07:40.000 --> 00:07:47.440 GT was less than half the price of its big brother GTX but offered anywhere 00:07:44.560 --> 00:07:52.400 from 70 to 90% of the performance depending on the application. And this 00:07:49.440 --> 00:07:56.319 was in a single slot form factor. It was such a clear winner in terms of value 00:07:54.240 --> 00:08:00.639 that NVIDIA has never offered anything similar since. Now I already declared 00:07:58.639 --> 00:08:05.360 GeForce 8 to be S tier, but it's worth saying it again. S tier, which is how 00:08:03.919 --> 00:08:09.919 you'll feel and look in our underwear from ltstore.com. 00:08:07.039 --> 00:08:14.720 Take it from me, I would know. After a wild previous generation, 00:08:11.919 --> 00:08:19.120 GeForce 9 was a bit of a rest on the old laurels time for NVIDIA. We went 00:08:17.039 --> 00:08:24.240 straight from the largest generational performance jump to the smallest one. 00:08:22.000 --> 00:08:31.120 The G92 core that was at the heart of the 9800 GTX, I mean, it was no slouch. 00:08:27.680 --> 00:08:33.039 It handily dispatched the ATIHD 3870, 00:08:31.120 --> 00:08:39.680 but it was based on the same Tesla architecture and in fact based on the 00:08:36.000 --> 00:08:42.719 same die as the 8800 GT, just shrunk for 00:08:39.680 --> 00:08:45.040 a smaller 55nm process node, which was 00:08:42.719 --> 00:08:49.440 good for a healthy clock speed bump, but only about a 10% performance increase 00:08:47.200 --> 00:08:55.519 versus last gen, unless of course you sprang for the dual GPU 9800 GX2 00:08:53.200 --> 00:08:59.440 sandwich edition. I guess we also got PCIe Gen 2 support across the entire 00:08:57.600 --> 00:09:04.000 range, but it would be a while before cards, and it wouldn't be these ones, 00:09:01.680 --> 00:09:08.880 managed to saturate that connection. I'm going with C tier. What was next? 00:09:07.360 --> 00:09:14.959 Right. Nothing. All five cards from the GeForce 00:09:12.080 --> 00:09:19.519 100 generation, yes, it existed, were OEM only. We've seen recent examples of 00:09:17.680 --> 00:09:24.480 GPUs that were sold exclusively to system integrators. The GTX 1065 gig is 00:09:22.480 --> 00:09:29.680 one that comes to mind, but an entire generation NVIDIA. The good news is we 00:09:27.519 --> 00:09:34.560 didn't miss out on anything because every 100 series card was rebadged from 00:09:32.240 --> 00:09:41.040 either the previous or the following generation. You get a big fat D tier. 00:09:37.920 --> 00:09:43.600 But ooh, this is new. NVIDIA's Tesla 00:09:41.040 --> 00:09:50.080 architecture is back again with the 200 series, but the GTX goes before the 00:09:46.640 --> 00:09:52.320 numbers now. Ooh. 00:09:50.080 --> 00:09:58.160 In all seriousness though, the flagship GTX 280 and its mid-gen refreshed 285 00:09:56.160 --> 00:10:03.600 used what NVIDIA called their second generation Tesla architecture on TSMC's 00:10:00.880 --> 00:10:09.519 65 and 55 nanmter process nodes respectively. And they were hefty mamas 00:10:06.720 --> 00:10:14.800 topping over 1 billion transistors for the first time ever. This was also the 00:10:12.320 --> 00:10:20.160 first and only time that we saw NVIDIA use a whopping 512-bit memory bus. 00:10:17.600 --> 00:10:26.560 granting gamers a very respectable 60ish% performance bump. And while the 00:10:23.680 --> 00:10:31.279 ATI HD4870 managed to take the frames per dollar crown, NVIDIA's top tier was 00:10:28.880 --> 00:10:36.480 in a different league altogether and still hovering around that $500 flagship 00:10:33.839 --> 00:10:42.240 price point. SLI had also matured significantly by this point, allowing 00:10:38.079 --> 00:10:44.560 gamers to drive highresolution 2560x600 00:10:42.240 --> 00:10:49.200 displays. And it was even beneficial for lower resolutions with scaling that 00:10:46.079 --> 00:10:51.360 could reach 70 to 80%. That means that 00:10:49.200 --> 00:10:54.959 two mid-range cards was a great value option compared to a single high-end 00:10:53.040 --> 00:11:02.399 one. I personally went that route with two GTX 260s. I'm going A tier. As for 00:10:59.200 --> 00:11:04.959 the GTX 300 series, 00:11:02.399 --> 00:11:10.320 but this again, another generation of five cards that was OEM only. thrilling 00:11:08.160 --> 00:11:16.640 except that no, it's actually worse than last time because they were all rebadged 00:11:12.959 --> 00:11:19.680 from the previous generation. Let's say 00:11:16.640 --> 00:11:22.480 F tier for forgettable. This next one 00:11:19.680 --> 00:11:27.680 wasn't. The first DirectX11 capable generation arrived 5 months later than 00:11:24.560 --> 00:11:30.480 planned, but with over 50% performance 00:11:27.680 --> 00:11:35.519 gains relative to the GTX 280, and it managed to edge out ATI's HD 5870. So 00:11:33.600 --> 00:11:40.320 clearly all those CUDA cores were worth waiting for. GPUs still use thousands of 00:11:38.399 --> 00:11:43.839 these computing cores for everything from rendering scenery to drawing 00:11:42.240 --> 00:11:49.600 character models to calculating lighting and shadows in a game environment. And 00:11:46.079 --> 00:11:53.440 all this performance came cheap. Okay, 00:11:49.600 --> 00:11:55.680 not cheap, but at an even $500. The GTX 00:11:53.440 --> 00:12:02.399 480 was priced identically to the flagship FX GPU from 7 years prior, 00:11:58.640 --> 00:12:04.640 making it $120 cheaper in today dollars 00:12:02.399 --> 00:12:11.120 if you account for inflation. Oh yeah, and it also cost $150 then dollars less 00:12:08.160 --> 00:12:15.279 than the GTX 280 that it replaced. It wasn't all rainbows and sparkles, 00:12:12.560 --> 00:12:20.880 though. It was also blazing inferno. The GTX furnace, uh, barbecue card, and 00:12:18.800 --> 00:12:25.279 thermy jokes were so common that even AMD joined in on the mockery. In terms 00:12:23.519 --> 00:12:29.360 of tech though, we got two big new arrivals. First, thanks to NVIDIA's 00:12:27.440 --> 00:12:33.839 acquisition of AIA, who up until this point had made, I think, PowerPoint 00:12:31.360 --> 00:12:38.560 presentations and promises. No, sorry. Uh, a weird standalone physics card. Um, 00:12:36.560 --> 00:12:43.279 well, yeah, that technology runs on the GPU now. And second was support for 00:12:40.880 --> 00:12:48.480 tessillation, a technique used to add detail to surfaces that might otherwise 00:12:45.519 --> 00:12:53.200 be rendered flat. Unfortunately, it was still early days for Tessellation and it 00:12:50.320 --> 00:12:58.880 incurred a 20% hit to frames per second, making it one of those features that 00:12:56.000 --> 00:13:03.760 doesn't really run so much as walk on first generation hardware 00:13:01.360 --> 00:13:09.440 or TX, excuse me. Sorry. Uh, we're going to go B tier here. The GTX 500 series, 00:13:07.200 --> 00:13:14.240 though, was a solid tick in that it focused heavily on improving rather than 00:13:11.839 --> 00:13:17.760 revolutionizing NVIDIA's existing Fermy chip micro architecture. This is a 00:13:16.160 --> 00:13:22.399 strategy they borrowed from Intel that resulted in lower temperatures, noise, 00:13:20.079 --> 00:13:27.440 and power consumption and also less mockery. Performance gains in the 20% 00:13:25.040 --> 00:13:31.839 range were a little on the low end, but NVIDIA did widen the gap compared to 00:13:29.440 --> 00:13:37.120 AMD's best, and pricing stayed the same at 500 bucks for the top-end single GPU 00:13:34.399 --> 00:13:42.000 card. Notably, GeForce 500 saw the reintroduction of the TI moniker, which 00:13:39.600 --> 00:13:47.519 hadn't been used since 2002. We're still not sure how to pronounce it. GTX 560 00:13:44.320 --> 00:13:49.360 Tai Ti. But we are sure that at the time 00:13:47.519 --> 00:13:53.040 cards with this designation represented incredible performance for the price. 00:13:51.600 --> 00:13:57.519 Fun fact, if you've ever had the misfortune, by the way, of buying a fake 00:13:54.880 --> 00:14:01.440 GPU from a site like Wish or AliExpress, you probably got a card from this 00:13:59.040 --> 00:14:06.000 generation with a modified BIOS. I'm going to go A tier. Now, one of the most 00:14:03.839 --> 00:14:10.639 notable things about our next generation GTX 600 is that NVIDIA's Kepler 00:14:08.720 --> 00:14:16.800 architecture was the first time they tried to pass off a mid-tier GPU as a 00:14:13.839 --> 00:14:22.959 flagship. See, here's the thing. The GTX 680 uses a GK 104 chip, but the flagship 00:14:20.720 --> 00:14:27.519 die of each generation almost always ends in a zero. And then the higher that 00:14:25.279 --> 00:14:31.680 last number goes, the smaller the die and the lower end the part. What that 00:14:29.920 --> 00:14:37.279 means is whether it was for cost concerns or an utter lack of competition 00:14:33.920 --> 00:14:38.800 from AMD, the GTX 680 that we got was a 00:14:37.279 --> 00:14:43.600 lesser version of what it could have been. But if you were expecting it to 00:14:40.800 --> 00:14:49.360 come with a discount, think again. To be fair, overall performance was up 20% and 00:14:47.199 --> 00:14:53.920 it was an enormous leap in performance per watt thanks to a risky strategy of 00:14:51.920 --> 00:14:58.000 dramatically shrinking the process node and significantly altering the 00:14:55.440 --> 00:15:01.600 architecture in one generation. And this is the point where NVIDIA's strategy of 00:14:59.519 --> 00:15:06.399 differentiating based on features starts to really make its mark with the 600 00:15:04.160 --> 00:15:12.720 series. Then we got some basics like support for PCI Express Gen 3 and 00:15:08.959 --> 00:15:15.120 DisplayPort 1.2, but we also got GPU 00:15:12.720 --> 00:15:20.880 boost, Envank hardware accelerated video encoding, G-Sync, less demanding FXA and 00:15:18.480 --> 00:15:24.959 TXA anti-aliasing, and support for the Vulcan API. And that's not all. A 00:15:23.600 --> 00:15:30.720 personal favorite of mine was the introduction of Gamestream. RIP NVIDIA's 00:15:28.560 --> 00:15:37.120 endto-end solution for streaming games from your GeForce powered PC to a Shield 00:15:33.279 --> 00:15:39.360 handheld device. So cool. It wasn't all 00:15:37.120 --> 00:15:43.680 roses, though. Driver issues that caused stuttering with adaptive VSYNC were 00:15:40.959 --> 00:15:49.199 widely reported. And confusingly, the total number of SKUs ballooned from 15 00:15:46.560 --> 00:15:54.959 to 27. And that's without factoring in the superc clocks and super super clocks 00:15:51.759 --> 00:15:57.120 of the world. Making matters worse, AMD 00:15:54.959 --> 00:16:00.880 sayanara ATI branding took its first performance crown in some time with the 00:15:58.959 --> 00:16:04.800 HD7970. I think we can go B tier, but I also 00:16:03.040 --> 00:16:10.240 think I'm being kind of generous here. GTX 700, on the other hand, can be best 00:16:07.120 --> 00:16:12.240 summarized as more Kepler. Due to the 00:16:10.240 --> 00:16:16.320 slowdown of Moore's law, NVIDIA was stuck with the same 28 nmter 00:16:14.160 --> 00:16:22.399 manufacturing process and Kepler architecture, but this time bigger, much 00:16:19.279 --> 00:16:24.399 bigger. That GK 110 chip that was 00:16:22.399 --> 00:16:28.480 exclusive to enterprise compute cards last time around finally made its first 00:16:26.560 --> 00:16:34.320 appearance for consumers in the form of the GTX Titan. And boy was that thing 00:16:31.120 --> 00:16:36.240 ever interesting. Depending on who you 00:16:34.320 --> 00:16:41.120 asked at the time, NVIDIA would tell you that Titan was intended for gamers or 00:16:38.560 --> 00:16:45.920 for professionals or for somewhere in between. And it did have two key 00:16:43.199 --> 00:16:51.839 advantages over GeForce GPUs. namely, it had a whopping 6 gigabytes of high-speed 00:16:48.639 --> 00:16:53.120 GDDDR5 memory and eight times better 00:16:51.839 --> 00:16:57.839 double precision floatingoint performance. But then in subsequent 00:16:55.360 --> 00:17:02.000 iterations of these Titan cards, those advantages would ultimately disappear. 00:16:59.759 --> 00:17:06.079 And in retrospect, I think the Titan project was simply a way for NVIDIA to 00:17:03.759 --> 00:17:12.880 test gamers tolerance for much higher prices. $1,000 for a GPU. We said, "Oh, 00:17:10.640 --> 00:17:18.000 young Lionus, you ain't seen nothing yet." Anywh who, the definitely for 00:17:16.160 --> 00:17:23.439 gamers real flagships for this generation were the GTX 780, which 00:17:20.319 --> 00:17:27.039 managed a 25% increase over last gen for 00:17:23.439 --> 00:17:29.120 $150 more, and the 780 Ti, which came 00:17:27.039 --> 00:17:33.679 only 6 months later and managed a similar performance jump for only a 00:17:31.200 --> 00:17:38.559 marginal increase in price. Meanwhile, AMD's code name Hawaii cards were hot 00:17:36.720 --> 00:17:42.240 and loud with the bigger issue being that they were plagued by driver issues 00:17:40.480 --> 00:17:46.880 and wouldn't really catch up in terms of performance for a couple more years. 7 00:17:45.200 --> 00:17:54.000 series then comes away with the score the 6 series should have gotten a tier. 00:17:50.960 --> 00:17:57.280 As for GTX 800 series, 00:17:54.000 --> 00:18:00.559 once again, it did exist, but instead of 00:17:57.280 --> 00:18:02.640 more Kepler, the theme was less Kepler 00:18:00.559 --> 00:18:08.400 in your laptop. Dt tier. Now, there were plans actually 00:18:05.840 --> 00:18:12.640 for a desktop line of 800 series cards apparently, but they were intended to be 00:18:10.400 --> 00:18:19.120 based on NVIDIA's new Maxwell architecture, which would instead make 00:18:15.280 --> 00:18:21.440 its debut as the GTX 900 series. Maxwell 00:18:19.120 --> 00:18:25.919 is notable for a number of reasons. It was the second time that NVIDIA passed a 00:18:23.440 --> 00:18:31.520 mid-tier GPU off as a flagship. It was the third desktop generation that was 00:18:28.160 --> 00:18:33.840 stuck on TSMC's 28 nanmter process. 00:18:31.520 --> 00:18:38.160 Moore's law was looking well and truly dead at this time. And it was the last 00:18:36.400 --> 00:18:43.280 time that we would see a new generation launch on a close enough to yearly 00:18:40.880 --> 00:18:51.360 cycle. Impressively though, Maxwell offered a 30% performance increase for 00:18:46.000 --> 00:18:54.640 wait less. Yes, the 980 was a $550 card 00:18:51.360 --> 00:18:56.799 at launch, $100 less than the 780. On 00:18:54.640 --> 00:19:02.160 top of that, it was AMD's turn to phone it in with a rebrand of their 290X to 00:18:59.200 --> 00:19:06.480 390X, leaving NVIDIA essentially alone at the top of the performance charts. It 00:19:04.400 --> 00:19:13.200 was also so efficient that for the first time ever, NVIDIA used the same desktop 00:19:09.440 --> 00:19:14.799 GPU dies in 900 series laptops. They had 00:19:13.200 --> 00:19:19.120 some interesting ideas for how to use their performance leadership. One was 00:19:16.799 --> 00:19:23.679 dynamic super resolution, which was possible to do before, but simplified 00:19:21.679 --> 00:19:28.080 and allowed you to render your game at a higher than native resolution, then 00:19:25.600 --> 00:19:32.480 downsample it for a smoother appearance. Another was to push gaming on 00:19:29.840 --> 00:19:38.720 highresolution monitors and TVs. Less demanding MFAA or multi-frame 00:19:34.559 --> 00:19:40.720 anti-aliasing and HDMI 2.0 made 4K 60 Hz 00:19:38.720 --> 00:19:46.320 gaming a major marketing point for NVIDIA. But with that said, even the 00:19:43.600 --> 00:19:51.760 properly high-end 980 Ti that came as a $750 midcycle refresh struggled to 00:19:49.280 --> 00:19:56.480 deliver a playable 4K experience in AAA games. I mean, to be clear, these were 00:19:54.000 --> 00:20:00.799 fast cards and shockingly efficient considering that the transistors were 00:19:57.919 --> 00:20:05.039 the same size as Kepler, but NVIDIA set our expectations maybe a little bit too 00:20:02.799 --> 00:20:09.760 high. Notable too from this generation was the introduction of VXGI, even if 00:20:07.840 --> 00:20:14.080 Voxal global illumination would ultimately give way to path-raced RTX 00:20:12.080 --> 00:20:18.240 lighting, which we'll get to later. And finally, it's impossible to talk about 00:20:16.080 --> 00:20:23.600 this generation without bringing up the 3 1/2 gig GTX 970 debacle for which 00:20:21.600 --> 00:20:28.240 NVIDIA eventually had to settle a class action lawsuit. The cards did have 4 00:20:26.000 --> 00:20:34.240 gigs of RAM as advertised. It's just that only 7/8 of it was actually 00:20:30.720 --> 00:20:36.000 accessible at full speed. Whoops. Yeah, 00:20:34.240 --> 00:20:42.000 you know what? It's still A tier in my book. Though maybe not as A tier as the 00:20:39.360 --> 00:20:48.080 GTX 10 series. We finally got a new manufacturing node, guys. Let's go. 00:20:45.039 --> 00:20:51.679 TSMC's 16 nmter process meant a healthy 00:20:48.080 --> 00:20:53.840 bump in well, everything. At launch, 00:20:51.679 --> 00:20:57.440 gamers could pick up a top tier card. Sort of. They did that thing again where 00:20:55.440 --> 00:21:02.960 they withheld the biggest die for a TI refresh with a very respectable 60% 00:21:00.640 --> 00:21:07.840 increase in performance for 600 bucks while drawing less power and producing 00:21:05.200 --> 00:21:12.640 less heat. 10 series also gave us our first taste of GPU Boost 3.0, which 00:21:11.120 --> 00:21:17.440 could best be described as self-overclocking, allowing the GPU to 00:21:14.960 --> 00:21:21.760 scale its clock speed well beyond base clock depending on available power and 00:21:19.280 --> 00:21:27.520 thermal headroom. For their part, AMD's flagship RX 480 was the value king at 00:21:25.600 --> 00:21:32.400 the time, but thanks to NVIDIA's investments in GPU encoding and the 00:21:29.679 --> 00:21:39.039 explosion of online game streaming, it was well overshadowed by the mid-range 00:21:35.280 --> 00:21:41.280 GTX 1060. Which one, you ask? Ah, good 00:21:39.039 --> 00:21:46.159 question. In another example of NVIDIA using extremely misleading product 00:21:43.200 --> 00:21:50.320 names, the 3 gig and 6 gig version of the 1060 have performance differences 00:21:48.159 --> 00:21:55.679 that go way beyond their frame buffer size. For shame 00:21:53.120 --> 00:22:00.559 three and four-way SLI were dropped, as was SLI altogether on mid-range cards, 00:21:58.400 --> 00:22:06.320 but NVIDIA did introduce a high bandwidth bridge as a bit of a swan song 00:22:03.360 --> 00:22:10.640 to two-way SLI. Speaking of bandwidth, the new DisplayPort 1.4 4 standard also 00:22:08.880 --> 00:22:15.760 continued the trend of catering towards high resolution, high refresh rate, and 00:22:13.120 --> 00:22:21.840 shiny new high dynamic range monitors. 4K 120 Hz HDR, anyone? I mean, not at 00:22:19.360 --> 00:22:25.440 that time, but it was so close. We could smell it. I think I'm going to go S tier 00:22:23.840 --> 00:22:28.960 here. It's just too bad that in hindsight, it turns out that what I 00:22:27.039 --> 00:22:35.760 smelled was the wet fart that was around the next corner. Alongside the 00:22:31.200 --> 00:22:39.280 confusingly named RTX 20/GTX16 00:22:35.760 --> 00:22:41.360 series came ray tracing. I don't know if 00:22:39.280 --> 00:22:46.240 I can remember the last time gamers paid so much for so few frames in return. 00:22:44.559 --> 00:22:50.400 Remember that really cool marble demo that NVIDIA put together to show off the 00:22:47.840 --> 00:22:55.039 capabilities of these new cards. Yeah, that ran at 25 frames per second at 00:22:53.280 --> 00:23:01.679 720p. And with these capabilities came a $100 00:22:57.919 --> 00:23:04.240 price hike. That's right, $700 for a, 00:23:01.679 --> 00:23:09.520 yes, they did it again, mid-tier GPU diet launch with the proper big TI one 00:23:06.960 --> 00:23:12.880 coming later. After the huge leap forward last generation, NVIDIA's 00:23:11.280 --> 00:23:18.480 performance gains in traditional games cooled to a more tepid 30-ish%. But I 00:23:15.679 --> 00:23:22.880 mean, hey, alongside our ray tracing RT course, we got Tensor course for cool 00:23:20.960 --> 00:23:27.039 machine learning accelerated features like background noise removal and deep 00:23:25.039 --> 00:23:31.840 learning super sampling, which kind of sucked at launch, but is like freaking 00:23:29.360 --> 00:23:36.320 amazing. Now, it's kind of wild seeing what a long game NVIDIA plays sometimes. 00:23:34.080 --> 00:23:40.240 Ray tracing is still making its way into the mainstream gaming space, but they 00:23:38.000 --> 00:23:44.000 were so sure it was the future of immersive, realistic in-game lighting 00:23:42.320 --> 00:23:48.799 that they went as far as to change their whole gaming brand. It's too bad the RTX 00:23:47.039 --> 00:23:53.760 20 series never ended up being capable of powering full-fledged rayraced games. 00:23:51.440 --> 00:24:00.640 Newer demos like RTX Racer and the new RTX Portal look amazing, but they only 00:23:57.520 --> 00:24:03.200 really run on newer cards. and RTX Remix 00:24:00.640 --> 00:24:08.159 looks set to rewrite the history books for people who skipped RTX 20 series. 00:24:06.320 --> 00:24:14.960 Sorry buddy, you're gonna have to be Ctier is what I would say about the 00:24:11.120 --> 00:24:18.559 RTX30 series in hindsight, but man, at 00:24:14.960 --> 00:24:20.159 launch this generation was lit. Gamers 00:24:18.559 --> 00:24:24.080 were so hyped to get their hands on these cards that our announcement 00:24:21.679 --> 00:24:29.360 coverage was the most viewed GPU announcement video we have ever done. 00:24:26.720 --> 00:24:34.400 The comments were so full of hope. More than double the total transistors and 00:24:31.440 --> 00:24:41.279 CUDA cores at the top end. The 3070 was outperforming the 2080 Ti at less than 00:24:37.600 --> 00:24:43.919 half the price. What? Too bad it was not 00:24:41.279 --> 00:24:48.559 to be. After the initial availability issues dampened our enthusiasm, the 00:24:46.080 --> 00:24:53.039 subsequent crypto mining boom completely dissolved. Secondary sellers, 00:24:50.960 --> 00:24:59.200 they're called scalpers, drove those coveted $500 RTX 3070s and 00:24:57.279 --> 00:25:05.360 every other card this generation to prices beyond double their MSRP. And 00:25:02.400 --> 00:25:09.120 there they stayed. I mean, on the plus side, AMD returned to competitive form 00:25:07.520 --> 00:25:14.000 and was able to offer similar performance at lower prices, unless it 00:25:11.919 --> 00:25:18.720 came to real-time ray traced lighting. And then, I mean, they too were shipping 00:25:15.919 --> 00:25:24.480 cards by the boatload to Cryptoarm. So, boo AMD as well. SLI also officially b 00:25:22.320 --> 00:25:30.559 us farewell at this stop in our journey, but we got to say hello to PCIe Gen 4, a 00:25:27.679 --> 00:25:35.679 stop gap 12 pin power connector, and support for direct storage, which allows 00:25:32.720 --> 00:25:39.279 your GPU and your SSD to talk directly, basically eliminating loading times, 00:25:37.600 --> 00:25:42.880 assuming any games ever show up with support for it. As for our tier 00:25:41.279 --> 00:25:47.120 assignment, I feel like we need to kind of split this one up. For anyone who got 00:25:45.200 --> 00:25:53.200 one of these legitimately ray tracing capable cards for MSRP, 30 series is an 00:25:49.919 --> 00:25:55.039 easy A. No, you know what? S tier. S 00:25:53.200 --> 00:26:00.159 tier. It delivered on the promises of last gen, but did it at attractive 00:25:56.960 --> 00:26:03.039 prices for everyone else though, I think 00:26:00.159 --> 00:26:08.559 a C is pretty generous. Oh, bringing us to the present day. Technically, 00:26:05.919 --> 00:26:12.559 NVIDIA's 40 series offerings aren't all out yet, but based on what we've seen so 00:26:10.960 --> 00:26:16.960 far, we can make some pretty safe guesses. In the coming months, we'll 00:26:14.559 --> 00:26:20.159 probably see some unlaunches. Just kidding, that was weird. Probably won't 00:26:18.480 --> 00:26:25.679 happen again. More likely, we'll see some yes launches, fleshing out the 00:26:22.080 --> 00:26:28.320 lineup with presumably RTX 4060 and 00:26:25.679 --> 00:26:32.240 4050s of some sort. And we also wouldn't be surprised to see some kind of new 00:26:30.080 --> 00:26:37.039 Halo card launch farther down the road, similar to how the 3090 Ti arrived a 00:26:34.880 --> 00:26:40.640 year and a half after its non-TI brother. We're going to have the ones 00:26:38.400 --> 00:26:46.000 that are out linked down below. Similar to last generation, we saw a pretty nice 00:26:43.200 --> 00:26:51.279 40% increase in performance. It's just that unlike last gen, prices didn't stay 00:26:49.200 --> 00:26:56.559 the same. They didn't even increase a little. They increased to a degree that 00:26:53.840 --> 00:27:04.400 I I don't think we've ever seen before. Like really, NVIDIA from $700 the 00:27:00.320 --> 00:27:06.159 previous two generations to $1,200. 00:27:04.400 --> 00:27:10.000 I think you guys got hooked on that pandemic Kool-Aid and haven't figured 00:27:08.000 --> 00:27:15.600 out that outside of the really wacky buyers who have always existed buying 00:27:12.640 --> 00:27:19.520 Titan Z's and the like. Gamers are not going to spend that kind of money to 00:27:17.039 --> 00:27:25.760 play video games. Hence, the shelves full of 4080s and 4070 Ti. For their 00:27:23.360 --> 00:27:31.120 part, AMD has retained a strong value proposition, but hasn't been able to 00:27:27.520 --> 00:27:33.520 match the insanity of the RTX 4090. and 00:27:31.120 --> 00:27:39.360 we got to meet a new power connector again. The 12volt high power connector 00:27:36.559 --> 00:27:42.799 has been a hot topic of conversation. It theoretically offers some cool benefits 00:27:41.039 --> 00:27:47.200 like allowing the GPU and power supply to communicate their power requirements 00:27:44.559 --> 00:27:50.640 and capabilities to each other. But uh with all the melted connectors and 00:27:48.960 --> 00:27:55.360 adapters that got me'd all over the internet, it's been a bit of a PR 00:27:52.480 --> 00:27:59.919 nightmare for NVIDIA who hit back saying that user error was the problem with 00:27:57.919 --> 00:28:03.279 pretty much everything that happened, which may actually be true. Please do 00:28:02.000 --> 00:28:09.120 make sure you plug things in all the way. In summary, if we're going by the 00:28:05.840 --> 00:28:13.679 kilogram, 40 series offers a fine value, 00:28:09.120 --> 00:28:15.679 but we aren't. So, Btier at best, and 00:28:13.679 --> 00:28:20.640 I'm generously baking in some future price drops once NVIDIA regains their 00:28:17.600 --> 00:28:23.120 senses. All that remains then now is for 00:28:20.640 --> 00:28:26.559 you to come at me, bro. Seriously, we couldn't get into the nitty-gritty of 00:28:24.559 --> 00:28:30.320 performance or talk about every new feature for all these generations. So, 00:28:28.320 --> 00:28:34.000 let me know on our forum or in the comments down below if you agree or 00:28:32.320 --> 00:28:37.679 disagree with our reasoning. What's your tier list and how does this video rank 00:28:36.000 --> 00:28:42.159 for you? If you guys enjoyed this video, you'll also probably like the full 00:28:39.279 --> 00:28:47.120 review we did of the RTX 470 Ti. That is unless you bought one, in which case you 00:28:44.000 --> 00:28:47.120 might not enjoy that review that