10 YEARS of NVIDIA Video Cards Compared!

Linus Tech Tips ·Linus Tech Tips ·2018-05-06 · 1,974 words · ~9 min read
Floatplane YouTube

Transcript

JSON SRT VTT 187
0:00 Now, as most of you probably know, NVIDIA acquired our company back in
0:04 early Q2 of 2016. So, you can imagine
0:07 how upset our Emerald overlords were when we decided to test 10 years worth
0:13 of AMD video cards before giving the
0:16 green team the same treatment. Well, we
0:19 finally put out the fire in our server room and just in time to bring you the
0:25 other side of the GPU performance coin.
0:28 Welcome to NVIDIA video cards through
0:32 the
0:42 ages. Rockat's Plus Force FX features
0:46 pressure sensitive QWES as keys which
0:50 can be configured in three different modes for multiple functions. Check it
0:54 out now at the link below. This is going to be a long video, so let's get right
0:59 into it. We used a 5960X test bench with
1:02 64 gigs of DDR4 RAM to isolate the
1:05 performance of the graphics cards. And we used the same benchmarks as AMD GPUs
1:11 through the ages for consistency. First up is NVIDIA's ancient flagship, the
1:18 8800 GTX. This card was freaking
1:21 revolutionary. It brought us an all new architecture called Tesla in the biggest
1:27 chip we had ever seen in a GPU code name
1:32 G80. It was built on the same 90mm
1:35 process node as its predecessor G71, but
1:39 was over twice as big with almost 2 and
1:44 1/2 times as many transistors devoted to
1:47 gaming performance, enabling new DirectX
1:50 10 features like SM 4.0 0 and a reduction in API overhead. CUDA, which
1:55 is still very much in use today and enabled developers to run generalpurpose
1:59 code on the graphics chip and the ability to run three cards at once in
2:05 SLI, which isn't to say that its standalone performance wasn't impressive
2:10 in its own right. G80 with its 768 megs
2:14 of GDDR3 RAM on a wide 384-bit bus
2:19 smacked around even two of NVIDIA's
2:22 previous flagships in SLI. And all of
2:25 that for a measly $100 more, which almost makes me feel better about the
2:29 cash grab 8800 Ultra that they released 6 months later for $230 more that was
2:35 basically the same thing with a fancy plastic shroud on it. The GeForce 8800
2:41 GT wasn't strictly speaking a new
2:45 flagship, but rather a weird Tesla
2:48 architecture refresh. It was the first built by NVIDIA on TSMC's shiny new 65
2:55 nanometer process. This node shrink allowed NVIDIA to drop power consumption
3:00 by 60 watts, reduce the die size, and
3:04 actually bump up the transistor count at the same time. a modest increase that
3:10 you guessed it. No, actually you probably didn't. The 8800 GT was on par
3:16 or slower than its predecessor almost
3:20 across the board due to its narrower memory bus and 16 fewer shader units.
3:27 Where people got excited was the price.
3:30 8800 GT was less than half the price of
3:34 8800 GTX. It was single slot, enabling
3:38 MATX users to run to an SLI with additional expansion cards, and it also
3:43 featured an onchip display engine.
3:46 Finally, it was the first consumer GPU with a PCI Express 2.0 interface, which
3:52 still to this day is barely saturated by
3:55 flagship graphics cards. The last flagship silicon before NVIDIA
4:00 transitioned to the GTX insert three numbers and maybe a couple letters here
4:05 nomenclature was G92, the beating heart
4:08 of the 9800 GTX Plus, a skew that
4:12 lowered NVIDIA's cost and bumped up performance over the architecturally
4:17 identical 9800 GTX by shrinking from 65
4:22 to 55 nm and boosting up the clock
4:25 speed. Now we get to the GTX 280. Yes,
4:30 we know it actually came out before that last one I was talking about, but deal
4:34 with it. So, forget everything we said earlier about 55 nanmter. The 280 built
4:40 on the mature 65nm process had a massive
4:45 576 mm die with 1.4 billion transistors.
4:50 And to put that insanity in context,
4:53 AMD's closest competitor, the 4870 had a
4:57 measly 959 million. And further driving up
5:01 complexity, it shipped with a minimum of 1 Gigabyte of graphics memory on a huge
5:08 512bit bus, something NVIDIA hasn't done
5:11 before or since on a consumer card. The
5:15 good news is that it absolutely crushed their previous generation efforts, even
5:20 delivering ample performance for stereoscopic 3D gaming, a big push for
5:25 NVIDIA at that time. We'll skip over the 300 series because all of these were
5:30 just crappy rebrands for notebooks and Best Buy PCs, meaning that the next real
5:35 GPU in line is the GTX 480 based on the
5:39 then all-new and now infamous Fermy
5:42 architecture. This was NVIDIA's first
5:46 swing at DirectX11 support, which is still widely used today. And this is
5:51 becoming a bit of a common theme. But while the new 40nm process allowed it to
5:56 be smaller than the GTX 280, it was still huge with a 529 millimeter die, a
6:04 wamping 3 billion transistors, and
6:08 support for DDDR5 memory, which again is
6:12 still widely used today. Interesting
6:15 fact, NVIDIA apparently thought it performed so well not just in gaming but
6:21 also in compute which was a big focus for Fermy that the largest configuration
6:27 of GF100 silicon never actually made its
6:30 way into a shipping product. Or maybe that was due to power constraints. Cuz
6:36 uh yeah, even with the theoretical
6:39 efficiency advantage of a smaller node, the 480 was such a barbecue that NVIDIA
6:45 provided a convenient grill on the face of the card to cook your eggs. Our next
6:50 contestant is the GTX 580, and it
6:54 featured an optimized and fully enabled Fermy dubbed GF 110. We've got most of
7:00 the usual improvements here. more VRAMm, higher clock speeds, and a slightly
7:06 smaller die, but nothing as exciting as video card cooking, which we mean in a
7:11 good way. 580 performed noticeably better than 480 while also managing
7:16 lower temperatures, noise, and power consumption. It was kind of like the GTX
7:23 480S, if that makes sense to the Apple fans out there. In the seventh corner of
7:28 our showdown, this must be a weird looking ring. We have the GTX 680. Ah,
7:33 yes. Goodbye Fermy. Hello Kepler. A lot
7:37 of things changed here. 4K 60 Hz output
7:40 was enabled via DisplayPort 1.2. PCI
7:45 Express 3.0 burst onto the scene with double the bandwidth of version two. The
7:50 die size shrunk way down to 294 mm. And
7:56 the GDDR5 got kicked up a notch to 6 GHz
8:00 thanks to a newly redesigned memory controller. Finally, we got it. Wait a
8:06 minute. We dropped down to a 256 bit
8:10 bus. Hold on a sec. Look at that code
8:13 name. NVIDIA top tier chips usually end
8:17 in a zero. Well, what they did was they
8:20 used AMD's complacency and a new 28nmter
8:23 process along with some dynamic power tricks called GPU Boost to crank up
8:29 performance so much that they gave us a
8:33 stepdown chip as a replacement for a top
8:36 tier one to save a buck and still manage to make it look like an upgrade. Not
8:42 that they passed any of those savings to the consumer. It was priced at $500,
8:47 same as the GTX 580. Though, I guess at least we got hardware H.264 encoding for
8:53 game stream and Shadow Play, G-Sync variable refresh rate technology, and
8:58 support for the Vulcan API. Now, the
9:01 next logical stop in our journey might seem like GTX 780, but in February 2013,
9:08 NVIDIA kicked convention and logic in the head with a Titan card based on the
9:16 chip that might have been GTX 680 in a
9:20 parallel universe where AMD had been keeping up, but instead it turned into
9:26 780 and then 780 Ti. So, we're going to
9:30 go with that since it was kind of the final form of the GK 110. Now, not much
9:37 changed product or feature-wise here.
9:40 Big Kepler was a much bigger, much
9:44 better 680 with a 90% bigger die, more
9:48 than double the transistors at a whopping 7 billion, and a $699 price tag
9:53 to go with it. Ouch. At least though,
9:57 this amounted to a sizable bump in performance over last generation's 680
10:04 in basically every application. In another unconventional move, NVIDIA
10:09 completely skipped over 800 series and went straight to 900. So, even though
10:15 the 980 and the 980 Ti were 9 months
10:18 apart, we're going to bunch them together cuz apparently that's what
10:22 everybody's doing. So, the 900 series was actually based on the same 28nm
10:27 process as Kepler. Shrinking silicon got
10:31 a lot harder in the early 2010s, but 2
10:34 years later, we were due for some kind of performance improvement. So, NVIDIA's
10:39 engineers brought us Maxwell. These GPUs
10:43 feature an integrated ARM CPU, which according to NVIDIA provides more
10:48 independence from a given systems primary CPU. The 980, again a step down
10:55 chip in big kid clothes, performs a little better than the Kepler GTX 780
11:00 Ti, which is an impressive feat. But the
11:04 big story is efficiency. It boasts an 85
11:08 watt lower TDP. The 980 Ti was a bit of
11:12 a different story, though. It rocked 8 billion transistors, an unheard of 601
11:17 square millimeter die, NVIDIA's biggest ever, 6 gigs of VRAM, and all the API
11:23 support you could ever want, except maybe DirectX12, depending on who you
11:29 ask. While Maxwell was marketed as fully
11:32 DirectX12 capable, the developer of Ashes of the Singularity, Oxide Games,
11:37 found that Maxwell cards performed very poorly with Async compute enabled, which
11:41 is kind of a big deal. So, all of this amounted to an astonishing performance
11:46 improvement considering the lack of process node change, and HDMI 2.0
11:51 support was a nice touch. Finally, we arrive at the modern day.
11:57 The GTX 1080 and 1080 Ti based on the
12:00 Pascal architecture. 10 series GPUs have
12:04 brought a lot to the table, including GDDR 5X at the top end, which increased
12:10 memory speeds to up to 10 GHz. Multiple
12:14 DisplayPort 1.4 four ports for 8K and
12:18 high refresh 4K monitor support, high
12:21 bandwidth SLI bridges, not to mention the elimination of 3-way and 4-way SLI
12:28 for gaming, and the smallest manufacturing node yet, 16 nanmter. 1080
12:35 launched in May of 2016 with 7.2 billion
12:38 transistors. And this is becoming a pattern now, a small performance
12:42 improvement over the 980 Ti. Then 1080
12:45 Ti followed with a much more substantial
12:48 bump. One of the biggest stories here though was really compute. Pascal was
12:54 designed to be a compute powerhouse with
12:57 the top tier GP100 chip going actually
13:01 unused in gaming products. So 1080 Ti
13:05 and its Titan analog were instead based
13:09 on GP 102, a new in between tier.
13:16 The ingenious EWS 360AP is an enterprise
13:20 class dualband access point that's ideal for large offices, schools, and
13:25 hospitality guest Wi-Fi access. It features AC speeds with business class
13:31 features and operates in autonomous or managed modes. It's got a 3x3 antenna
13:37 design and a high-owered radio and comes with a free management software called
13:41 Easy Master. You can manage anywhere from one access point to thousands. And
13:46 there are no per device license fees. So
13:50 that means no annual subscription, which is great for keeping the total cost of
13:54 ownership down. Check it and other ingenious products out over at their
13:59 website through the link in the video
14:02 description. Thanks for watching NVIDIA GPUs through the ages. If this video
14:07 sucked, you know what to do. But if it was awesome, get subscribed, hit the like button, or check out the link to
14:11 where to buy the stuff we featured in the video description. Also linked down
14:14 there is our merch store, which has cool shirts like this one, and our community
14:18 forum, which you should totally join.