WEBVTT 00:00:00.080 --> 00:00:07.680 Now, as most of you probably know, NVIDIA acquired our company back in 00:00:04.319 --> 00:00:10.000 early Q2 of 2016. So, you can imagine 00:00:07.680 --> 00:00:16.240 how upset our Emerald overlords were when we decided to test 10 years worth 00:00:13.040 --> 00:00:19.920 of AMD video cards before giving the 00:00:16.240 --> 00:00:22.480 green team the same treatment. Well, we 00:00:19.920 --> 00:00:28.560 finally put out the fire in our server room and just in time to bring you the 00:00:25.439 --> 00:00:32.079 other side of the GPU performance coin. 00:00:28.560 --> 00:00:34.100 Welcome to NVIDIA video cards through 00:00:32.079 --> 00:00:39.069 the 00:00:42.120 --> 00:00:50.239 ages. Rockat's Plus Force FX features 00:00:46.239 --> 00:00:52.320 pressure sensitive QWES as keys which 00:00:50.239 --> 00:00:57.360 can be configured in three different modes for multiple functions. Check it 00:00:54.640 --> 00:01:02.559 out now at the link below. This is going to be a long video, so let's get right 00:00:59.520 --> 00:01:05.600 into it. We used a 5960X test bench with 00:01:02.559 --> 00:01:08.479 64 gigs of DDR4 RAM to isolate the 00:01:05.600 --> 00:01:14.880 performance of the graphics cards. And we used the same benchmarks as AMD GPUs 00:01:11.920 --> 00:01:21.840 through the ages for consistency. First up is NVIDIA's ancient flagship, the 00:01:18.080 --> 00:01:24.560 8800 GTX. This card was freaking 00:01:21.840 --> 00:01:32.280 revolutionary. It brought us an all new architecture called Tesla in the biggest 00:01:27.759 --> 00:01:35.920 chip we had ever seen in a GPU code name 00:01:32.280 --> 00:01:39.759 G80. It was built on the same 90mm 00:01:35.920 --> 00:01:44.000 process node as its predecessor G71, but 00:01:39.759 --> 00:01:47.200 was over twice as big with almost 2 and 00:01:44.000 --> 00:01:50.240 1/2 times as many transistors devoted to 00:01:47.200 --> 00:01:52.560 gaming performance, enabling new DirectX 00:01:50.240 --> 00:01:57.520 10 features like SM 4.0 0 and a reduction in API overhead. CUDA, which 00:01:55.680 --> 00:02:02.719 is still very much in use today and enabled developers to run generalpurpose 00:01:59.840 --> 00:02:08.479 code on the graphics chip and the ability to run three cards at once in 00:02:05.960 --> 00:02:14.560 SLI, which isn't to say that its standalone performance wasn't impressive 00:02:10.640 --> 00:02:19.120 in its own right. G80 with its 768 megs 00:02:14.560 --> 00:02:22.319 of GDDR3 RAM on a wide 384-bit bus 00:02:19.120 --> 00:02:25.360 smacked around even two of NVIDIA's 00:02:22.319 --> 00:02:28.080 previous flagships in SLI. And all of 00:02:25.360 --> 00:02:32.560 that for a measly $100 more, which almost makes me feel better about the 00:02:29.840 --> 00:02:37.760 cash grab 8800 Ultra that they released 6 months later for $230 more that was 00:02:35.519 --> 00:02:45.360 basically the same thing with a fancy plastic shroud on it. The GeForce 8800 00:02:41.800 --> 00:02:48.560 GT wasn't strictly speaking a new 00:02:45.360 --> 00:02:51.519 flagship, but rather a weird Tesla 00:02:48.560 --> 00:02:58.160 architecture refresh. It was the first built by NVIDIA on TSMC's shiny new 65 00:02:55.200 --> 00:03:04.640 nanometer process. This node shrink allowed NVIDIA to drop power consumption 00:03:00.480 --> 00:03:07.360 by 60 watts, reduce the die size, and 00:03:04.640 --> 00:03:12.640 actually bump up the transistor count at the same time. a modest increase that 00:03:10.720 --> 00:03:20.959 you guessed it. No, actually you probably didn't. The 8800 GT was on par 00:03:16.800 --> 00:03:23.440 or slower than its predecessor almost 00:03:20.959 --> 00:03:30.239 across the board due to its narrower memory bus and 16 fewer shader units. 00:03:27.040 --> 00:03:34.159 Where people got excited was the price. 00:03:30.239 --> 00:03:38.400 8800 GT was less than half the price of 00:03:34.159 --> 00:03:41.280 8800 GTX. It was single slot, enabling 00:03:38.400 --> 00:03:46.640 MATX users to run to an SLI with additional expansion cards, and it also 00:03:43.599 --> 00:03:49.440 featured an onchip display engine. 00:03:46.640 --> 00:03:55.760 Finally, it was the first consumer GPU with a PCI Express 2.0 interface, which 00:03:52.640 --> 00:03:58.319 still to this day is barely saturated by 00:03:55.760 --> 00:04:03.120 flagship graphics cards. The last flagship silicon before NVIDIA 00:04:00.480 --> 00:04:08.560 transitioned to the GTX insert three numbers and maybe a couple letters here 00:04:05.000 --> 00:04:12.159 nomenclature was G92, the beating heart 00:04:08.560 --> 00:04:15.040 of the 9800 GTX Plus, a skew that 00:04:12.159 --> 00:04:22.479 lowered NVIDIA's cost and bumped up performance over the architecturally 00:04:17.680 --> 00:04:25.919 identical 9800 GTX by shrinking from 65 00:04:22.479 --> 00:04:30.560 to 55 nm and boosting up the clock 00:04:25.919 --> 00:04:32.800 speed. Now we get to the GTX 280. Yes, 00:04:30.560 --> 00:04:36.639 we know it actually came out before that last one I was talking about, but deal 00:04:34.880 --> 00:04:45.240 with it. So, forget everything we said earlier about 55 nanmter. The 280 built 00:04:40.160 --> 00:04:50.400 on the mature 65nm process had a massive 00:04:45.240 --> 00:04:53.680 576 mm die with 1.4 billion transistors. 00:04:50.400 --> 00:04:57.120 And to put that insanity in context, 00:04:53.680 --> 00:04:58.360 AMD's closest competitor, the 4870 had a 00:04:57.120 --> 00:05:04.240 measly 959 million. And further driving up 00:05:01.520 --> 00:05:11.520 complexity, it shipped with a minimum of 1 Gigabyte of graphics memory on a huge 00:05:08.440 --> 00:05:15.280 512bit bus, something NVIDIA hasn't done 00:05:11.520 --> 00:05:17.919 before or since on a consumer card. The 00:05:15.280 --> 00:05:22.479 good news is that it absolutely crushed their previous generation efforts, even 00:05:20.400 --> 00:05:28.160 delivering ample performance for stereoscopic 3D gaming, a big push for 00:05:25.360 --> 00:05:32.479 NVIDIA at that time. We'll skip over the 300 series because all of these were 00:05:30.240 --> 00:05:39.440 just crappy rebrands for notebooks and Best Buy PCs, meaning that the next real 00:05:35.199 --> 00:05:42.880 GPU in line is the GTX 480 based on the 00:05:39.440 --> 00:05:46.000 then all-new and now infamous Fermy 00:05:42.880 --> 00:05:48.720 architecture. This was NVIDIA's first 00:05:46.000 --> 00:05:53.440 swing at DirectX11 support, which is still widely used today. And this is 00:05:51.680 --> 00:05:59.440 becoming a bit of a common theme. But while the new 40nm process allowed it to 00:05:56.479 --> 00:06:08.800 be smaller than the GTX 280, it was still huge with a 529 millimeter die, a 00:06:04.759 --> 00:06:12.560 wamping 3 billion transistors, and 00:06:08.800 --> 00:06:15.759 support for DDDR5 memory, which again is 00:06:12.560 --> 00:06:18.160 still widely used today. Interesting 00:06:15.759 --> 00:06:24.319 fact, NVIDIA apparently thought it performed so well not just in gaming but 00:06:21.600 --> 00:06:30.880 also in compute which was a big focus for Fermy that the largest configuration 00:06:27.440 --> 00:06:33.680 of GF100 silicon never actually made its 00:06:30.880 --> 00:06:39.840 way into a shipping product. Or maybe that was due to power constraints. Cuz 00:06:36.120 --> 00:06:42.000 uh yeah, even with the theoretical 00:06:39.840 --> 00:06:47.360 efficiency advantage of a smaller node, the 480 was such a barbecue that NVIDIA 00:06:45.120 --> 00:06:54.000 provided a convenient grill on the face of the card to cook your eggs. Our next 00:06:50.880 --> 00:06:56.960 contestant is the GTX 580, and it 00:06:54.000 --> 00:07:03.680 featured an optimized and fully enabled Fermy dubbed GF 110. We've got most of 00:07:00.960 --> 00:07:09.280 the usual improvements here. more VRAMm, higher clock speeds, and a slightly 00:07:06.319 --> 00:07:14.240 smaller die, but nothing as exciting as video card cooking, which we mean in a 00:07:11.599 --> 00:07:19.639 good way. 580 performed noticeably better than 480 while also managing 00:07:16.720 --> 00:07:25.599 lower temperatures, noise, and power consumption. It was kind of like the GTX 00:07:23.000 --> 00:07:29.919 480S, if that makes sense to the Apple fans out there. In the seventh corner of 00:07:28.319 --> 00:07:37.440 our showdown, this must be a weird looking ring. We have the GTX 680. Ah, 00:07:33.759 --> 00:07:40.800 yes. Goodbye Fermy. Hello Kepler. A lot 00:07:37.440 --> 00:07:45.039 of things changed here. 4K 60 Hz output 00:07:40.800 --> 00:07:47.759 was enabled via DisplayPort 1.2. PCI 00:07:45.039 --> 00:07:56.479 Express 3.0 burst onto the scene with double the bandwidth of version two. The 00:07:50.400 --> 00:08:00.479 die size shrunk way down to 294 mm. And 00:07:56.479 --> 00:08:02.879 the GDDR5 got kicked up a notch to 6 GHz 00:08:00.479 --> 00:08:10.120 thanks to a newly redesigned memory controller. Finally, we got it. Wait a 00:08:06.160 --> 00:08:13.400 minute. We dropped down to a 256 bit 00:08:10.120 --> 00:08:17.039 bus. Hold on a sec. Look at that code 00:08:13.400 --> 00:08:20.160 name. NVIDIA top tier chips usually end 00:08:17.039 --> 00:08:23.680 in a zero. Well, what they did was they 00:08:20.160 --> 00:08:26.400 used AMD's complacency and a new 28nmter 00:08:23.680 --> 00:08:33.279 process along with some dynamic power tricks called GPU Boost to crank up 00:08:29.759 --> 00:08:36.560 performance so much that they gave us a 00:08:33.279 --> 00:08:39.519 stepdown chip as a replacement for a top 00:08:36.560 --> 00:08:44.159 tier one to save a buck and still manage to make it look like an upgrade. Not 00:08:42.240 --> 00:08:50.000 that they passed any of those savings to the consumer. It was priced at $500, 00:08:47.279 --> 00:08:56.160 same as the GTX 580. Though, I guess at least we got hardware H.264 encoding for 00:08:53.360 --> 00:09:01.839 game stream and Shadow Play, G-Sync variable refresh rate technology, and 00:08:58.640 --> 00:09:04.000 support for the Vulcan API. Now, the 00:09:01.839 --> 00:09:11.839 next logical stop in our journey might seem like GTX 780, but in February 2013, 00:09:08.959 --> 00:09:20.240 NVIDIA kicked convention and logic in the head with a Titan card based on the 00:09:16.399 --> 00:09:22.959 chip that might have been GTX 680 in a 00:09:20.240 --> 00:09:30.800 parallel universe where AMD had been keeping up, but instead it turned into 00:09:26.560 --> 00:09:32.640 780 and then 780 Ti. So, we're going to 00:09:30.800 --> 00:09:40.800 go with that since it was kind of the final form of the GK 110. Now, not much 00:09:37.360 --> 00:09:44.320 changed product or feature-wise here. 00:09:40.800 --> 00:09:48.240 Big Kepler was a much bigger, much 00:09:44.320 --> 00:09:50.080 better 680 with a 90% bigger die, more 00:09:48.240 --> 00:09:57.279 than double the transistors at a whopping 7 billion, and a $699 price tag 00:09:53.920 --> 00:09:59.760 to go with it. Ouch. At least though, 00:09:57.279 --> 00:10:07.040 this amounted to a sizable bump in performance over last generation's 680 00:10:04.560 --> 00:10:12.399 in basically every application. In another unconventional move, NVIDIA 00:10:09.519 --> 00:10:18.480 completely skipped over 800 series and went straight to 900. So, even though 00:10:15.200 --> 00:10:20.079 the 980 and the 980 Ti were 9 months 00:10:18.480 --> 00:10:25.040 apart, we're going to bunch them together cuz apparently that's what 00:10:22.480 --> 00:10:31.440 everybody's doing. So, the 900 series was actually based on the same 28nm 00:10:27.839 --> 00:10:34.560 process as Kepler. Shrinking silicon got 00:10:31.440 --> 00:10:36.959 a lot harder in the early 2010s, but 2 00:10:34.560 --> 00:10:43.680 years later, we were due for some kind of performance improvement. So, NVIDIA's 00:10:39.680 --> 00:10:46.480 engineers brought us Maxwell. These GPUs 00:10:43.680 --> 00:10:50.480 feature an integrated ARM CPU, which according to NVIDIA provides more 00:10:48.480 --> 00:10:58.079 independence from a given systems primary CPU. The 980, again a step down 00:10:55.360 --> 00:11:04.000 chip in big kid clothes, performs a little better than the Kepler GTX 780 00:11:00.720 --> 00:11:08.320 Ti, which is an impressive feat. But the 00:11:04.000 --> 00:11:12.079 big story is efficiency. It boasts an 85 00:11:08.320 --> 00:11:14.320 watt lower TDP. The 980 Ti was a bit of 00:11:12.079 --> 00:11:19.839 a different story, though. It rocked 8 billion transistors, an unheard of 601 00:11:17.519 --> 00:11:26.640 square millimeter die, NVIDIA's biggest ever, 6 gigs of VRAM, and all the API 00:11:23.920 --> 00:11:32.240 support you could ever want, except maybe DirectX12, depending on who you 00:11:29.120 --> 00:11:34.480 ask. While Maxwell was marketed as fully 00:11:32.240 --> 00:11:39.279 DirectX12 capable, the developer of Ashes of the Singularity, Oxide Games, 00:11:37.120 --> 00:11:44.000 found that Maxwell cards performed very poorly with Async compute enabled, which 00:11:41.839 --> 00:11:49.040 is kind of a big deal. So, all of this amounted to an astonishing performance 00:11:46.079 --> 00:11:54.320 improvement considering the lack of process node change, and HDMI 2.0 00:11:51.760 --> 00:12:00.959 support was a nice touch. Finally, we arrive at the modern day. 00:11:57.040 --> 00:12:04.800 The GTX 1080 and 1080 Ti based on the 00:12:00.959 --> 00:12:06.800 Pascal architecture. 10 series GPUs have 00:12:04.800 --> 00:12:14.399 brought a lot to the table, including GDDR 5X at the top end, which increased 00:12:10.560 --> 00:12:18.480 memory speeds to up to 10 GHz. Multiple 00:12:14.399 --> 00:12:21.519 DisplayPort 1.4 four ports for 8K and 00:12:18.480 --> 00:12:24.399 high refresh 4K monitor support, high 00:12:21.519 --> 00:12:30.959 bandwidth SLI bridges, not to mention the elimination of 3-way and 4-way SLI 00:12:28.800 --> 00:12:38.800 for gaming, and the smallest manufacturing node yet, 16 nanmter. 1080 00:12:35.200 --> 00:12:40.959 launched in May of 2016 with 7.2 billion 00:12:38.800 --> 00:12:45.519 transistors. And this is becoming a pattern now, a small performance 00:12:42.480 --> 00:12:48.560 improvement over the 980 Ti. Then 1080 00:12:45.519 --> 00:12:50.959 Ti followed with a much more substantial 00:12:48.560 --> 00:12:57.279 bump. One of the biggest stories here though was really compute. Pascal was 00:12:54.079 --> 00:13:01.000 designed to be a compute powerhouse with 00:12:57.279 --> 00:13:05.839 the top tier GP100 chip going actually 00:13:01.000 --> 00:13:09.279 unused in gaming products. So 1080 Ti 00:13:05.839 --> 00:13:15.600 and its Titan analog were instead based 00:13:09.279 --> 00:13:15.600 on GP 102, a new in between tier. 00:13:16.880 --> 00:13:23.279 The ingenious EWS 360AP is an enterprise 00:13:20.720 --> 00:13:28.320 class dualband access point that's ideal for large offices, schools, and 00:13:25.680 --> 00:13:34.160 hospitality guest Wi-Fi access. It features AC speeds with business class 00:13:31.200 --> 00:13:39.440 features and operates in autonomous or managed modes. It's got a 3x3 antenna 00:13:37.120 --> 00:13:43.519 design and a high-owered radio and comes with a free management software called 00:13:41.519 --> 00:13:50.160 Easy Master. You can manage anywhere from one access point to thousands. And 00:13:46.560 --> 00:13:52.399 there are no per device license fees. So 00:13:50.160 --> 00:13:57.360 that means no annual subscription, which is great for keeping the total cost of 00:13:54.800 --> 00:14:02.680 ownership down. Check it and other ingenious products out over at their 00:13:59.120 --> 00:14:05.120 website through the link in the video 00:14:02.680 --> 00:14:08.880 description. Thanks for watching NVIDIA GPUs through the ages. If this video 00:14:07.279 --> 00:14:12.880 sucked, you know what to do. But if it was awesome, get subscribed, hit the like button, or check out the link to 00:14:11.199 --> 00:14:16.160 where to buy the stuff we featured in the video description. Also linked down 00:14:14.800 --> 00:14:23.399 there is our merch store, which has cool shirts like this one, and our community 00:14:18.399 --> 00:14:23.399 forum, which you should totally join.