NVIDIA’s Gamescom 2018 keynote just wrapped up, and as many have been expecting since it was announced last month, NVIDIA is getting ready to launch their next generation of GeForce hardware. Announced at the event and going on sale starting September 20th is NVIDIA’s GeForce RTX 20 series, which is succeeding the current Pascal-powered GeForce GTX 10 series. Based on NVIDIA’s new Turing GPU architecture and built on TSMC’s 12nm “FFN” process, NVIDIA has lofty goals, looking to drive an entire paradigm shift in how games are rendered and how PC video cards are evaluated. CEO Jensen Huang has called Turing NVIDIA’s most important GPU architecture since 2006’s Tesla GPU architecture (G80 GPU), and from a features standpoint it’s clear that he’s not overstating matters.

As is traditionally the case, the first cards out of the NVIDIA stable are the high-end cards. But in a rather sizable break from tradition we’re not only going to get the x80 and x70 cards at launch, but also the x80 Ti card as well. Meaning the GeForce RTX 2080 Ti, RTX 2080, and RTX 2070 will all be hitting the streets within a month of each other. NVIDIA’s product stack is remaining unchanged here, so RTX 2080 Ti remains their flagship card, while RTX 2080 is their high-end card, and then RTX 2070 the slightly cheaper card to entice enthusiasts without breaking the bank.

All three cards will be launching over the next two months. First off will be the RTX 2080 Ti and RTX 2080, which will launch September 20th. The RTX 2080 Ti will start at $999 for partner cards, while the RTX 2080 will start at $699. Meanwhile the RTX 2070 will launch at some point in October, with partner cards starting at $499. On a historical basis, all of these prices are higher than the last generation by anywhere between $120 and $300. Meanwhile NVIDIA’s own reference-quality Founders Edition cards are once again back, and those will carry a $100 to $200 premium over the baseline pricing.

Unfortunately, NVIDIA is already taking pre-orders here, so consumers are essentially required to make a “blind buy” if they want to snag a card from the first batch. NVIDIA has offered surprisingly little information on performance and we’d suggest waiting for trustworthy third-party reviews (i.e. us), however I have to admit that I don’t imagine there’s going to be much stock available by the time reviews hit the streets.

NVIDIA GeForce Specification Comparison
  RTX 2080 Ti RTX 2080 RTX 2070 GTX 1080
CUDA Cores 4352 2944 2304 2560
Core Clock 1350MHz 1515MHz 1410MHz 1607MHz
Boost Clock 1545MHz 1710MHz 1620MHz 1733MHz
Memory Clock 14Gbps GDDR6 14Gbps GDDR6 14Gbps GDDR6 10Gbps GDDR5X
Memory Bus Width 352-bit 256-bit 256-bit 256-bit
VRAM 11GB 8GB 8GB 8GB
Single Precision Perf. 13.4 TFLOPs 10.1 TFLOPs 7.5 TFLOPs 8.9 TFLOPs
Tensor Perf. 440T OPs
(INT4)
? ? N/A
Ray Perf. 10 GRays/s 8 GRays/s 6 GRays/s N/A
"RTX-OPS" 78T 60T 45T N/A
TDP 250W 215W 175W 180W
GPU Big Turing Unnamed Turing Unnamed Turing GP104
Transistor Count 18.6B ? ? 7.2B
Architecture Turing Turing Turing Pascal
Manufacturing Process TSMC 12nm "FFN" TSMC 12nm "FFN" TSMC 12nm "FFN" TSMC 16nm
Launch Date 09/20/2018 09/20/2018 10/2018 05/27/2016
Launch Price MSRP: $999
Founders $1199
MSRP: $699
Founders $799
MSRP: $499
Founders $599
MSRP: $599
Founders $699

NVIDIA's Turing Architecture: RT & Tensor Cores

So what does Turing bring to the table? The marquee feature across the board is hybrid rendering, which combines ray tracing with traditional rasterization to exploit the strengths of both technologies. This announcement is essentially a continuation of NVIDIA’s RTX announcement from earlier this year, so if you thought that announcement was a little sparse, well then here is the rest of the story.

The big change here is that NVIDIA is going to be including even more ray tracing hardware with Turing in order to offer faster and more efficient hardware ray tracing acceleration. New to the Turing architecture is what NVIDIA is calling an RT core, the underpinnings of which we aren’t fully informed on at this time, but serve as dedicated ray tracing processors. These processor blocks accelerate both ray-triangle intersection checks and bounding volume hierarchy (BVH) manipulation, the latter being a very popular data structure for storing objects for ray tracing.

NVIDIA is stating that the fastest GeForce RTX part can cast 10 Billion (Giga) rays per second, which compared to the unaccelerated Pascal is a 25x improvement in ray tracing performance.

The Turing architecture also carries over the tensor cores from Volta, and indeed these have even been enhanced over Volta. The tensor cores are an important aspect of multiple NVIDIA initiatives. Along with speeding up ray tracing itself, NVIDIA’s other tool in their Turing bag of tricks is to reduce the amount of rays required in a scene by using AI denoising to clean up an image, which is something the tensor cores excel at. Of course that’s not the only feature tensor cores are for – NVIDIA’s entire AI/neural networking empire is all but built on them – so while not a primary focus for the Gamescom crowd, this also confirms that NVIDIA’s most powerful neural networking hardware will be coming to a wider range of GPUs.

Looking at hybrid rendering in general, it’s interesting that despite these individual speed-ups, NVIDIA’s overall performance promises aren’t quite as extreme. All told, the company is promising a 6x performance boost versus Pascal, and this doesn’t specify against which parts. Time will tell if even this is a realistic assessment, as even with the RT cores, ray tracing in general is still quite the resource hog.

As for gaming matters in particular, the benefits of hybrid rendering are potentially significant, but it’s going to depend heavily on how developers choose to use it. From performance standpoint I’m not sure there’s much to say here, and that’s because ray tracing & hybrid rendering are ultimately features to improve rendering quality, not improve the performance of today’s algorithms. Granted, if you tried to do ray tracing on today’s GPUs it would be extremely slow – and Turing an incredible speedup as a result – but no one uses slow path tracing systems on current hardware for this reason. So hybrid rendering is instead about replacing the approximations and hacks of current rasterization technology with more accurate rendering methods. In other words, less “faking it” and more “making it.”

Those quality benefits, in turn, are typically clustered around lighting, shadows, and reflections. All three features are inherently based on the properties of light, which in simplistic terms moves as a ray, and which up to now various algorithms have been faking the work involved or “pre-baking” scenes in advance. And while current algorithms are quite good, they still aren’t close to accurate. So there is clear room for improvement.

NVIDIA for their part is particularly throwing around global illumination, which is one of the harder tasks. However there are other lighting methods that benefit as well, not to mention reflections and shadows of those lit objects. And truthfully this is where words are a poor tool; it’s difficult to describe how a ray traced shadow looks better than a fake shadow with PCSS, or real-time lighting over pre-baked lighting. Which is why NVIDIA, the video card company, is going to be pushing the visual aspects of all of this harder than ever.

Overall then, hybrid rendering is the lynchpin feature of the GeForce RTX 20 series. Going by their Gamescom and SIGGRAPH presentations, it’s clear that NVIDIA has invested heavily into the field, and that they have bet the success of the GeForce brand over the coming years on this technology. RT cores and tensor cores are semi-fixed function hardware; they can’t be used for rasterization, and the transistors allocated to them are transistors that could have been dedicated to more rasterization hardware otherwise. So NVIDIA has made an incredibly significant move here in terms of opportunity cost by going the hybrid rendering route rather than building a bigger Pascal.

As a result, NVIDIA is attempting a paradigm shift in consumer rendering, one that we’ve really only see before with the introduction of pixel and vertex shaders (DX8 & DX9 era tech) all the way back in 2001 & 2002. Which is why Microsoft’s DirectX Raytracing (DXR) initiative is so important, as are NVIDIA’s other developer and consumer initiatives. NVIDIA needs to sell consumers and developers alike on this vision of mixing rasterization with ray tracing to provide better image quality. And more so than that, they need to ease developers into the idea of working with more specialized, fixed function units as Moore’s Law continues to slow down and fixed function hardware becomes a means to achieve greater efficiency.

NVIDIA hasn’t bet the farm on hybrid rendering, but they’ve never attempted to move the market in this fashion. So if it seems like NVIDIA is hyper-focused on hybrid rendering and ray tracing, that’s because they are. It’s their vision of the future, and now they need to get everyone else on board.

Turing SM: Dedicated INT Cores, Unified Cache, Variable Rate Shading

Alongside the dedicated RT and tensor cores, the Turing architecture Streaming Multiprocessor (SM) itself is also learning some new tricks. In particular here, it’s inheriting one of Volta’s more novel changes, which saw the Integer cores separated out into their own blocks, as opposed to being a facet of the Floating Point CUDA cores. The advantage here – at least as much as we saw in Volta – is that it speeds up address generation and Fused Multiply Add (FMA) performance, though as with a lot of aspects of Turing, there’s likely more to it (and what it can be used for) than we’re seeing today.

The Turing SM also includes what NVIDIA is calling a “unified cache architecture.” As I’m still awaiting official SM diagrams from NVIDIA, it’s not clear if this is the same kind of unification we saw with Volta – where the L1 cache was merged with shared memory – or if NVIDIA has gone one step further. At any rate NVIDIA is saying that it offers twice the bandwidth of the “previous generation” which is unclear if NVIDIA means Pascal or Volta (with the latter being more likely).

Finally, also tucked away in the SIGGRAPH Turing press release is the mention of support for variable rate shading. This is a relatively young and upcoming graphics rendering technique that there's limited information about (especially as to how exactly NVIDIA is implementing it). But at a very high level it sounds like the next generation of NVIDIA's multi-res shading technology, which allows developers to render different areas of a screen at various effective resolutions, in order to concentrate quality (and rendering time) in to the areas where it's the most beneficial.

Feeding the Beast: GDDR6 Support

As the memory used by GPUs is developed by outside companies, there are no big secrets here. The JEDEC and its big 3 members Samsung, SK Hynix, and Micron, have all been developing GDDR6 memory as the successor to both GDDR5 and GDDR5X, and NVIDIA ha confirmed that Turing will support it. Depending on the manufacturer, first-generation GDDR6 is generally promoted as offering up to 16Gbps per pin of memory bandwidth, which is 2x that of NVIDIA’s late-generation GDDR5 cards, and 40% faster than NVIDIA’s most recent GDDR5X cards.

GPU Memory Math: GDDR6 vs. HBM2 vs. GDDR5X
  NVIDIA GeForce RTX 2080 Ti
(GDDR6)
NVIDIA GeForce RTX 2080
(GDDR6)
NVIDIA Titan V
(HBM2)
NVIDIA Titan Xp
 
NVIDIA GeForce GTX 1080 Ti NVIDIA GeForce GTX 1080
Total Capacity 11 GB 8 GB 12 GB 12 GB 11 GB 8 GB
B/W Per Pin 14 Gb/s 1.7 Gb/s 11.4 Gbps 11 Gbps
Chip capacity 1 GB (8 Gb) 4 GB (32 Gb) 1 GB (8 Gb)
No. Chips/KGSDs 11 8 3 12 11 8
B/W Per Chip/Stack 56 GB/s 217.6 GB/s 45.6 GB/s 44 GB/s
Bus Width 352-bit 256-bit 3092-bit 384-bit 352-bit 256-bit
Total B/W 616 GB/s 448GB/s 652.8 GB/s 547.7 GB/s 484 GB/s 352 GB/s
DRAM Voltage 1.35 V 1.2 V (?) 1.35 V

Relative to GDDR5X, GDDR6 is not quite as big of a step up as some past memory generations, as many of GDDR6’s innovations were already baked into GDDR5X. None the less, alongside HBM2 for very high end use cases, it is expected to become the backbone memory of the GPU industry. The principle changes here include lower operating voltages (1.35v), and internally the memory is now divided into two memory channels per chip. For a standard 32-bit wide chip then, this means a pair of 16-bit memory channels, for a total of 16 such channels on a 256-bit card. While this in turn means there is a very large number of channels, GPUs are also well-positioned to take advantage of it since they are massively parallel devices to begin with.

NVIDIA for their part has confirmed that the first GeForce RTX cards will run their GDDR6 at 14Gbps, which happens to be the fastest speed grade offered by all of the Big 3 members. We know that NVIDIA is exclusively using Samsung's GDDR6 for their Quadro RTX cards – presumably because they need the density – however for the GeForce RTX cards the field should be open to all of the memory manufacturers. Though in the long run this leaves two avenues open to higher capacity cards: either moving up to 16Gb density chips, or going clamshell with the 8Gb chips they’re using now.

Odds & Ends: NVLink SLI, VirtualLink, & 8K HEVC

While this wasn’t mentioned in NVIDIA’s Gamescom presentation itself, NVIDIA’s GeForce 20 Series website confirms that SLI will once again be available for some high-end GeForce RTX cards. Specifically, both the RTX 2080 Ti and RTX 2080 will support SLI. Meanwhile the RTX 2070 will not support SLI; this being a departure from the 1070 which did offer it.

However the bigger aspect of that news is that NVIDIA’s proprietary cache coherent GPU interconnect, NVLink, will be coming to consumer cards. The GeForce GTX cards will be implementing SLI over NVLInk, with 2 NVLink channels running between each card. At a combined 50GB/sec of full-duplex bandwidth – meaning there’s 50GB of bandwidth available in each direction – this is a major upgrade over NVIDIA’s previous HB-SLI link. This is on top of NVLink’s other feature benefits, particularly cache coherence. And all of this comes at an important time, as inter-GPU bandwidth requirements keep rising with each generation.

Now the big question is whether this will reverse the ongoing decline of SLI, and at the moment I’m taking a somewhat pessimistic approach, but I’m eager to hear more from NVIDIA. 50GB/sec is a big improvement over HB-SLI, however it’s still only a fraction of the 448GB/sec (or more) of local memory bandwidth available to a GPU. So on its own it doesn’t fix the problems that have dogged multi-GPU rendering, either with AFR synchronization or effective workload splitting. In that respect it’s likely telling that NVIDIA doesn’t support NVLink SLI on the RTX 2070.

Meanwhile gamers something new to look forward to for VR, with the addition of VirtualLink support. The USB Type-C alternate mode was announced last month, and supports 15W+ of power, 10Gbps of USB 3.1 Gen 2 data, and 4 lanes of DisplayPort HBR3 video all over a single cable. In other words, it’s a DisplayPort 1.4 connection with extra data and power that is intended to allow a video card to directly drive a VR headset. The standard is backed by NVIDIA, AMD, Oculus, Valve, and Microsoft, so GeForce RTX cards will be the first of what we expect will ultimately be a number of products supporting the standard.

USB Type-C Alternate Modes
  VirtualLink DisplayPort
(4 Lanes)
DisplayPort
(2 Lanes)
Base USB-C
Video Bandwidth (Raw) 32.4Gbps 32.4Gbps 16.2Gbps N/A
USB 3.x Data Bandwidth 10Gbps N/A 10Gbps 10Gbps + 10Gbps
High Speed Lane Pairs 6 4
Max Power Mandatory: 15W
Optional: 27W
Optional: Up To 100W

Finally, while NVIDIA only briefly touched upon the subject, we do know that their video encoder block, NVENC, has been updated for Turing. The latest iteration of NVENC specifically adds support for 8K HEVC encoding. Meanwhile NVIDIA has also been able to further tune the quality of their encoder, allowing them to achieve similar quality as before with a 25% lower video bitrate.

Previewing GeForce RTX 2080 Ti
Comments Locked

223 Comments

View All Comments

  • Yojimbo - Tuesday, August 21, 2018 - link

    He never said 8800 GTX, he said 8800. There were a lot of 8800 GPUs and the 8800 Ultra launched for $830 in 2007, which is over 1,000 2018 dollars.
  • eddman - Tuesday, August 21, 2018 - link

    Then he'd be wrong if he meant Ultra, since 8800 GTX is the equivalent card to 2080. Also, the 8 series launched with 8800 GTX, not Ultra, which launched 6 months later simply because of lack of proper competition from AMD in order to maximize profits.
  • Yojimbo - Tuesday, August 21, 2018 - link

    He never said 2080, either, so why is her wrong? The 2080 doesn't cost $1,000. One can argue that the 2080 Ti is the 8800 Ultra coming out 6 months earlier, for which is actually deserves a price premium. As technology performance increases, the value of current technology goes down over time, so something coming out now is more valuable than something coming out 6 months from now.
  • eddman - Tuesday, August 21, 2018 - link

    ... because he replied to my comment where I compared 2080 to 1080. Isn't that obvious?

    No, 2080 Ti can be compared to 1080 Ti, 980 Ti, etc. and it is quite more expensive.
  • Yojimbo - Tuesday, August 21, 2018 - link

    I am not following your reasoning. You mentioned the 2080 and the 2080 Ti. You compared the 2080 to the 1080. Why does that preclude him from being able to compare the 2080 to the GTX 8800 and the 2080 Ti to the 8800 Ultra?

    The 2080 Ti can be compared to 1080 Ti but not to 8800 Ultra? Why? Because you said so? The 1080 Ti does not match his "new technology" thesis. If you're gonna refute his thesis you need to work with it, not just ignore it.

    As far as supporting his thesis, I think a better argument is the increase in die size of the GPUs in question and the fact that in NVIDIA's latest earnings conference call they projected that gross margins would go down next quarter, not up. That suggests that, assuming they sell a significant number of these new cards next quarter, these increased prices do seem to have a correlation to increased manufacturing costs, and not just "NVIDIA raising prices because they can". If they were raising prices "because they can" (which every company is trying to do!) then their margins should be increasing.
  • eddman - Tuesday, August 21, 2018 - link

    ... because at first I was comparing 2080 to 1080 and he replied to that.

    8800 Ultra was a massively overpriced card because of no competition. It didn't offer anything besides a bit higher clock speeds. It simply existed to make more money. If that new technology's cost was that crippling, nvidia would've priced 8800 GTX more.

    Thinking that these cards are this expensive ONLY because of their new technologies is naive. I don't doubt they could sell these at Pascal level prices and still make a lot of profit. They'd have simply recuperated their expenses a bit later.

    These new technologies might be playing a role in pricing but based on nvidia's track record it's safe to say that the biggest reason here is lack of competition.

    P.S. I have nothing against nvidia and might even upgrade to a 2070 from my 1060.
  • Yojimbo - Tuesday, August 21, 2018 - link

    "It simply existed to make more money."

    Products in general simply exist to make money. Of course someone can develop or champion something for another reason, but if it isn't fulfilling a desire in the market, in which case it can make money, then it's bound to fail.

    "Thinking that these cards are this expensive ONLY because of their new technologies is naive."

    No it's not, as shown by the gross margins argument I have made.

    "These new technologies might be playing a role in pricing but based on nvidia's track record it's safe to say that the biggest reason here is lack of competition."

    That simply isn't true. Based on the track record the prices have not gone up in relation despite NVIDIA's lack of competition in the PC gaming space. It's just something people repeat. In fact what has happened is a greater number of people have bought the higher priced cards than in previous generations. Why? Because the graphics cards have become more and more important to the performance of gaming systems in relation to the other components of the system. There also is less of a need to spend extra money on the system for uses of the systems besides gaming. Spreadsheets, word processors, compression tools, maintenance utilities, etc, used to tax
    the hardware relative to its maximum performance much more than they do now. Therefore that money saved on the other uses for the system can be re-allocated towards the graphics card to improve the gaming experience.
  • eddman - Tuesday, August 21, 2018 - link

    The point was that 8800 UItra was a mere cash grab. It didn't have to be that expensive or even exist, but nvidia did it since they knew AMD can't respond and that there are enough people to sell these cards to. Easy money. Nothing inherently wrong with that tough.

    Margins are expected to go down but it doesn't mean they couldn't have priced the cards lower. Yes, they made a lot of investments which had an effect on profit, but there is no reason to think that lower prices would've resulted in losses. They still would've made the money back but at a bit later date.

    Prices DID go up after fermi.
  • Yojimbo - Tuesday, August 21, 2018 - link

    "The point was that 8800 UItra was a mere cash grab. It didn't have to be that expensive or even exist, but nvidia did it since they knew AMD can't respond and that there are enough people to sell these cards to. Easy money. Nothing inherently wrong with that tough."

    Why do any graphics cards have to exist? Why can't NVIDIA raise chickens instead? Huang can be a long-haul trucker. The only way a product "has to be a certain price" is if a company can't possibly sell it for less without going out of business, I guess. If a company is pricing their products that way then they are in big, big trouble. So of course it was a "cash grab" priced to maximize profits. That's how products are priced! It's ubiquitous. It isn't special to the 8800 Ultra. And since it is ubiquitous the argument cannot possibly be used to refute a claim that the 8800 Ultra cost more because of new technology. It's entirely possible that the only reason people were willing to pay that much money for the card was because of its combination of performance and new technology. You have to realize that companies don't just set the price and then that's it. A company must always obey the market with their prices, otherwise they will lose money in comparison to obeying the market. It's the market that sets the price.

    "Margins are expected to go down but it doesn't mean they couldn't have priced the cards lower."

    Yes sure, margins could have gone down even more. Margins could even go to zero. They could sell the card at cost for a while before they tank their business. What's your point? We are arguing about whether they are "squeezing consumers" more not whether they are NOT "squeezing consumers" less!

    "Prices DID go up after fermi."

    No, the GTX 480 debuted at $499. 499 2010 dollars is 577 2018 dollars. Besides, AMD was still competitive during the Kepler days. The GTX 780 debuted at 649 2013 dollars, which is 702 2018 dollars. That's actually right where the GTX 2080 is debuting. Maxwell was when AMD was least competitive and the GTX 980 debuted at 549 2014 dollars which is 584 2018 dollars. So the price went down significantly, right in line with the GTX 480 debut price. This even though NVIDIA had much less competition with Maxwell and the GPU was more important to the total gaming performance of the system in 2014 compared to 2010.
  • Yojimbo - Tuesday, August 21, 2018 - link

    What I think is true is that with more AMD competition in the past prices dropped after after the debut. You can't use the Pascal generation for that comparison though, because it was a unique situation due to the crypto-currency craze and skyrocketing DRAM prices.

Log in

Don't have an account? Sign up now