Unintentionally, NVIDIA envisioned an overcast vision
The GeForce RTX 2000 series paved the way for segmentation. Strangely, the Greens may not necessarily do this, but the abuse of dominance does!
contents show
Gamescom 2018
As might be expected, NVIDIA laid out the pages for Gamescom 2018, so there was no surprise in this regard, but in other respects it was. The first: the flagship of the new series is not the GeForce GTX 1180, but the RTX 2080 Ti. The RTX mark is a reference to Ray tracing, practically an essential element of the new series. Let's start at the beginning.
The event was, of course, led by Jen-Hsun Huang, the show started with some slippage as usual. Much of the gunpowder has already been blown up at SIGGRAPH, so it didn’t come as a surprise to anyone that the Turing GPU that laid the foundations for the new series was born in the spirit of Ray tracing. Beam tracking is often described as the Holy Grail of computer graphic mapping, and NVIDIA’s first man should now (should) have convinced the audience that the technology was finally within reach. The process is indeed promising, with unprecedented realism in modeling the path of light, especially in reflections and shadows - the main problem is the insane performance greed of the implementation, although it must be acknowledged that it is at least excellent for parallelization, which is the strength of graphics processors. .
The past 8 Year 10
The first “CUDA GPU,” the GeForce 8800 GTX (which NVIDIA incorrectly demonstrated as a GTX 8800), came out in late 2006. According to Jen-Hsun Huang, barely two years after the release of the card, they started thinking about a much bigger change, as the development of the Turing chip just presented is the result of 10 years of work. If this is indeed the case, then it can be seen as a very bold move, as the rasterized graphics were still able to show spectacular progress after 2006 - just think of the Crysis introduced in 2008. An interesting draw was the mention of the G80 GPU (the heart of the 8800 GTX); the chip's biggest innovation was clearly the unified shader architecture, while Turing brings the future by accumulating fixed-function units, so we can say that there has indeed been a change of attitude.
RT
Turing runs computationally intensive Ray tracing on dedicated RT cores, with the most powerful solution at 10 gigays / s, or 10 billion beams per second - while the GeForce 1080 Ti offers just 1,21 gigars / s. , of course without RT cores. The progress is truly amazing at first glance, the four Volta GPUs in the DGX-1 can now be replaced with a Turing chip. RT cores are therefore dedicated Ray tracing processors in their interpretation, accelerating radius-triangle intersection checks in terms of their function.
Radiation tracking will definitely be supported by these games:
- Assasso Corsa Competizione (Simulation Games / 505 Games)
- Atomic Heart (Mundfish)
- Battlefield V (EA / DICE)
- Control Remedy (Entertainment / 505 Games)
- Enlisted Gaijin (Entertainment / Darkflow Software)
- Justice (NetEase)
- JX3 (Kingsoft)
- MechWarrior 5: Mercenaries (Piranha Games)
- Metro Exodus (4A Games)
- ProjectDH Nexon's (devCAT Studio)
- Shadow of the Tomb Raider (Square Enix / Eidos-Montreal / Crystal Dynamics / Nixxes)
Tensor
In contrast to RT cores, the existence of Tensor processors could not have been a novelty for fans, because the components optimized for the machine learning algorithm were already included in the previous architecture. Fans of the GeForce line will probably only be interested in the procedure called DLSS (Deep Learning Super Sampling); the Tensor cores compute a higher-resolution image from existing image information, and the result is used to improve image quality — perhaps even as a form of intelligent edge washing. Machine learning, by the way, offers really exciting opportunities (let’s take a look at the cutting-edge part of the presentation!), Only we doubt for ourselves that this would be a valid consideration for a hard-line player. Incidentally, the wide range of applications of Tensor cores is exemplified by their ability to filter out rays that would only unnecessarily load the GPU.
And these addresses will be interesting because of the DLSS:
- Ark: Survival (Evolved Studio Wildcard)
- Atomic Heart (Mundfish)
- Dauntless (Phoenix Labs)
- Final Fantasy XV (Square Enix)
- Fractured Lands (Unbroken Studios)
- Hitman 2 (IO Interactive / Warner Bros.)
- Islands of Nyne (Define Human Studios)
- Justice (NetEase)
- JX3 (Kingsoft)
- Mechwarrior 5: Mercenaries (Piranha Games)
- PlayerUnknown's Battlegrounds (PUBG Corp.)
- Remnant: the Ashes (Arc Games)
- Serious Sam 4: Planet Badass (Croteam / Devolver Digital)
- Shadow of the Tomb Raider (Square Enix / Eidos-Montreal / Crystal Dynamics / Nixxes)
- The Forge Arena (Freezing Raccoon Studios)
- We Happy Few (Compulsion Games / Gearbox)
CUDA
Streaming multiprocessors (SM) have carried forward one of Volta’s major innovations: they have disconnected the integer cores, a step that in principle speeds up Fused Multiply Add (FMA) operations and address generation. Another substantial development is the unified cache architecture, but in the absence of a block diagram, there are still question marks here. If Volta’s legacy lives on, L1 cache will be merged with shared memory. We have managed to find out that the bandwidth of the first level storage has doubled, the capacity has increased significantly and the access time has decreased significantly. L2 did not remain intact either: it increased from 3 MB to 6 MB. All of these changes are intended to provide Turing computing units with the right amount of work.
A little of everything
Unusually, NVIDIA was the first to listen to overall performance, so pre-orders can only go blind. The company promises a six-fold increase in performance against Pascal, of course in hybrid rendering. Speaking of which: the benefits of hybrid rendering are significant at first glance, but at the same time it depends heavily on how developers take advantage of the technology. In terms of speed, no progress is expected now because the procedure is primarily used to improve image quality. On today’s GPUs, beam tracking is capable of an invaluable pace, compared to which Turing promises tremendous acceleration, but that’s still enough to complement the raster graphics. This plus is expected to take the form of shadows, reflections, lights. Of course, there are already quite good algorithms in all three areas, but they are far from being as lifelike as the computer world.
Overall, hybrid rendering puts NVIDIA at risk, as the RT and Tensor cores are fixed-function hardware; they cannot be used for rasterization, but the transistor frame they digested could have been used in areas that would have accelerated traditional imaging. So far, we can witness a paradigm shift, with the Greens laying the groundwork for the future with the DirectX Raytracing (DXR) initiative, designed in close collaboration with Microsoft, at least according to them. We hope they are not mistaken, because otherwise they will be paddling the market into dangerous waters with this pull.
Feeding the Turing GPU
The paper form was brought by GDDR6 support. Progress after the GDDR5X is not earth-like, with a frequency of up to 16,0 GHz still providing enough bandwidth with a sufficiently wide data bus. This pace wasn’t even required by the first GeForce RTX cards (or at least they didn’t get it), NVIDIA pushed the limit at 14 Gbps.
Demos
There was no shortage of technical demonstrations, Ray tracing here and there has already hit its head. The first to show a Shadow of the Tomb Raider demo was where the programmers wanted to draw attention to themselves with lifelike shadows. This went well for the first scene (albeit with a highly artificial effect), less so for the second. It also turns out that the first step of ray tracing into real color will not always promise overwhelming production, but let’s move nicely in line.
Metro Exodus focused on global illumination. As we know, in the case of indirect illumination, the color of the pixel no longer depends only on the direct light source, but is also affected by the reflection of light on different surfaces. The result is perhaps more spectacular than Shadow of the Tomb Raider, but again, we feel that the developers are somewhat “attached” to it. Another delicate question arose in us: in our view, invisibly, most players would consider the image based on local illumination to be more beautiful, or at least more playable.
The really serious show was provided by Battlefield V. The alpha-level implementation of Ray Tracing has undoubtedly been a hit, but we are not convinced that the procedure will have a place in multiplayer mode. It is worth mentioning that the developers implemented the support in just two weeks.
The Three Musketeers
Eventually, the big unveiling took place, the new GeForce RTX series was introduced. The family currently has three members, these are RTX 2070, 2080 and 2080 Ti. Oddly enough, we only got foggy numbers from the raw tempo, but some progress is sure to happen even when the Tensor and RT cores are just milling because the Streaming Multiprocessor (SM) is also smart. What is certain is that the new cards have brought a massive price increase, as even the smallest model will be $ 499, the RTX 2080 will already hurt 699 and the top predator 999 will hurt the green belly. Founders Editions packs in an additional $ 100 for this, so for a surcharge, we get exactly 90 MHz higher clock speeds for all three types. The series can be pre-ordered now, but it will be available on September 20 at the earliest.
Unlike before, the Founder Edition also offers real, tangible benefits: 6 percent higher tempo at 10 watts. A common point, however, is the new cooling system: the two-fan monster is said to be able to operate significantly quieter than its predecessor. The new cooling, on the other hand, can suffer the rest of the PC, as this type of design tends to dissipate heat, meaning it doesn’t pull it out of the case as well as a closed system. Of course, it requires two slots, the length of the card is 10,5 inches. The RTX 2080 Ti is powered by a 13-phase power supply due to its not-so-modest power consumption and better overdrive, and covers power requirements with two eight-pin PCIe connectors. VGA can drive four displays, for which we have the following outputs: 3 DisplayPort and one HDMI.
Virtual Link
It was not yet clear at the Gamescom event what the situation is with multi-card mode. This was no accident because only the RTX 2080 Ti and RTX 2080 support SLI, the RTX 2070 no longer. The good news, though, is that HB-SLI will be replaced by NVLInk; progress can be made with a full duplex bandwidth of 50 GB / s. Perhaps more important than this is VirtualLink support. The recent standard is essentially a DisplayPort 1.4 connection for both power and data transfer; 15 watts and 10 Gbps USB 3.1 Gen 2. The standard is supported by NVIDIA, AMD, Oculus, Valve and Microsoft, and its main merit is that it allows virtual reality glasses to be connected to a computer on a single wire.
NVENC has also evolved, with support for 8K HEVC materials, and engineers say similar image quality at 25 percent lower bitrates.
Summary
So the graphics revolution could happen step by step, NVIDIA sees the future in hybrid rendering for now. This means that traditional raster graphics are only complemented by ray tracing (e.g. shadows, reflections, light effects). It can also be seen that this already comes at a very serious price, as the chip spun with Tensor and RT cores is brutally large (the top model is 754 mm²) on the one hand, and extremely complex on the other (consisting of 18,6 billion transistors). This means that with the present manufacturing technology (12 nm from TSMC), NVIDIA could only bring the technology to an acceptable price level by seriously sacrificing profits, which is clearly not in its interest. For this reason, we believe that Ray tracing will be just a premium feature in the current generation, a kind of PhysX 2.0 with the difference that only a narrow layer can now access it. Even if there are cheaper variants over time (RTX 2060, 2050), they will have to cut resources radically due to further cost reductions, which will almost certainly not jump to the minimum required. We can already give an example of the latter: in the Shadow of the Tomb Raider 1080p resolution, the GeForce RTX 2080 Ti controller was unable to run at a stable 60 fps, sometimes half of it was barely half. It can be seen from this that developers have to approach the technique very carefully, otherwise it could easily happen that the performance of the RTX 2070 will be insufficient.
The abuse of a dominant position in the introduction is (partly) reflected in prices. Although no one expected mass releases to appear for the above reasons, the GeForce RTX 2080 Ti’s $ 1 opening (Founder Edition) makes the Titan model unnecessary, as it previously moved at exactly that level. Perhaps even the most exciting for gamers, the RTX 199 is poisonous, appearing exactly 2070 percent more expensive than its predecessor - a price that is no longer far behind the GeForce GTX 31, which is pretty shocking. It looks like NVIDIA is only targeting ray tracking for players with thick wallets for the time being, which may have released their pandora box.
THE PROHARDWARE! according to his latest information, AMD may respond to Turing cards in the short term with a strong commentary pull. According to the portal, it is possible that Strange Brigade, Star Control: Origins, Assassin's Creed Odyssey, Resident Evil 2 (remake), Tom Clancy's The Division 2 and Doom: Eternal will include specific codes that have some unique give Radeon cards benefits; these will not necessarily show in performance, but may have different effects. If this proves to be the case, the times may come again when different manufacturers do not offer uniform graphic quality, i.e. we have to decide after individual consideration which games we want the best possible visual quality and then buy a product based on.
Source: HOC, AnandTech