Nvidia

Raytracing, tensor cores and AI: why do we need a video card from Nvidia for 96 thousand rubles

               

We understand what the company still showed and what PC players should expect in the near future.

What happened

After months of speculation, plums and rumors, the manufacturer of graphics processors Nvidia presented new video cards of the GeForce RTX 20xx family as part of the presentation at the exhibition Gamescom . The company announced the date of the performance in advance, and many immediately assumed that it would be about new products. However, in early June, the head of Nvidia assured that the next generation would have to wait for a long time, probably trying to support the demand for cards already on sale.

Models of this line are designed for a wider audience and work with games, unlike previously announced Titan V , Quadro GV100 and Quadro RTX for more professional tasks. Therefore, it is presented RTX 2070, RTX 2080 and the flagship RTX 2080 Ti should be considered a full-fledged development of the previous line GTX 10xx, which was waiting for PC gamers.

It is interesting that after more than a decade of use designations GTX, Nvidia abandoned it in favor of RTX – it means emphasis on ray tracing (or hybrid ray tracing) in real time. This technology has become the main innovation of future devices – this is a very resource-intensive process of accurate modeling of multiple reflections of light rays from various objects in the scene.

What is special about ray tracing?

First you need to understand what is so complicated in raytracing, how the new lighting processing differs from the already familiar one, and also why all this is necessary for ordinary players. To imagine how it works, simply look around the surrounding space, find the illuminated objects and trace where the light comes from. In fact, this is raytracing – “ray tracking”.

At the same time, if you put an obstacle between the light and the object or move the source itself (or even several), the shadows will also move, and sometimes even new ones will appear. The purpose of Nvidia technology is just to fully simulate this effect.

In words, raytracing may seem complicated, but in fact almost everyone has come across it at least once, even if not in games. Such effects are often used in modern films to create and enhance special effects, such as realistic reflections, refractions and shadows, although not always to the full. In this case, the shots taken with the camera are combined with objects created on the computer, which allows, if desired, drastically changing the final picture, creating the right feeling for the viewer.

The main obstacle to the use of raytracing in games for many years remained limited resources of computer hardware, which simply could not cope with the processing of frames in real time. Indeed, unlike cinema, where authors can pre-render the necessary frames for several dozen hours on special “farms” from professional video cards, in games for all this there is only a small fraction of a second.

Therefore, in computer graphics for a long time, a sufficiently fast technique of the so-called “rasterization” was used, allowing to display three-dimensional objects on a two-dimensional screen. In normal rasterization, objects in the scene are most often created from a special grid of triangles – polygons that intersect with each other (they can be of different sizes and shapes).


Rasterization

Different information is tied to such intersection points, including data on their position in space, color, textures and so-called “normals”. The latter are used for embossed texturing to make the surfaces of objects more voluminous and realistic. Very often, the shading effect changes as well, making the model look almost like a real one

After that, the video card converts the resulting models into pixels that the user sees on his screen. In addition, each of these pixels may be a different color, for which the above-mentioned data is used. Then “shading” is included in the work, with the help of which the pixels can receive additional effects and change color again depending on how the light falls on them.

Despite the simpler technology as compared to ray tracing, rasterization is still complex from the point of view of processing with iron. In the scene there can be a huge number of objects consisting of many hundreds of thousands of polygons, and pixels, if you take a resolution of 4K, on ​​the screen are about eight million. In addition, the image is updated from 30 (on consoles and weak PCs) to 60 and more (powerful PCs with expensive video cards) once a second


Demonstration of various techniques of conventional shadow rendering

Buffer helps to organize a constant stream of frames without delay, which also stores information about the state of pixels, which determines which of them should be displayed on the display, and which objects in the scene are behind and should remain hidden.

Raytracing is fundamentally different. In the real world, all visible objects are most often illuminated by several sources, but the light itself can be reflected several times from them even before it reaches the human eye, or to meet an obstacle, which causes a shadow. In addition, if there is something transparent or translucent, such as glass or water, the viewer will also see numerous light refractions and shadow overlaps


Raytracing


Tracing takes an in-game camera as an “eye” and tracks the path of each light beam through all the pixels on a two-dimensional display to the original 3D models. Thanks to this, developers can achieve better lighting quality and different lighting effects, which was not possible with more simple technologies. And although over the years of graphics development, many techniques appeared to fake honest effects, their potential was still very limited

How Nvidia works with raytracing

For its raytracing, Nvidia used several past developments of well-known figures from the gaming and film industry, including capturing reflections and refractions. For example, if the beam is reflected from surfaces several times or passes through several objects before reaching the light source, then the color and other information about the illumination from these objects will also affect the final appearance of the individual pixels.

In some cases, the use of ray tracing in games will not only improve the image, but will directly affect the gameplay. For example, the enemy can be seen earlier in the reflection and react in time. Probably, the technology will be used in stealth action, where more realistic lighting and complex shadows will help to better focus on locations full of opponents

The head of the company Jensen Huang at the last presentation said that the new graphics card line RTX 20xx will be the first devices for full ray tracing in real time. However, it is worth noting that the technology support will be in the past GTX 10xx family, although only at the initial level

The main difference of future video cards from previous models will be the presence of installed RT Cores on the board – several cores specifically designed to work with new algorithms and ensure better performance with raytracing enabled. Their exact number is not yet known, but the company still told about other features of the devices


The shadow appearing on the spacesuit becomes richer as the arm moves closer to the body.

The architecture of Turing, on which the new video cards are built, according to Nvidia, is a development of the ideas of a more professional Volta, designed to work with machine learning. Among other things, the devices have a new unified cache, which has twice the bandwidth compared to the previous generation, as well as a new streaming multiprocessor architecture that allows you to perform the calculation of integer expressions in parallel with performing floating point operations.

In addition, RT Cores themselves were enhanced by so-called tensor cores, designed for calculations of AI and neural networks. In short, this allows you to quickly identify possible errors in calculations or missing information during the ray tracing process, increasing the potential performance of the cards in this mode

In the new lineup, the number of so-called CUDA-cores – universal computing units of the video card has noticeably increased. In addition to basic indicators such as work frequencies, bandwidth and amount of memory, engineers also added a couple of new ones – a characteristic of Giga Rays / sec, reflecting the capabilities of each model when working with tracing, and RTX-OPS, which is the average number of all operations performed by the device, including ray tracing , shading and other tasks.

By all indications, the RTX models bypassed the old devices of the same price categories. Initially, representatives of Nvidia at the conference paid attention only to raytracing and demonstrations of his work, but after a couple of days the company clarified that even without taking into account the trace, the performance of new video cards is about one and a half times higher than that of the GTX 10xx models.

Judging by the published performance comparison of the RTX 2080 and GTX 1080, with a 4K resolution, the device still managed to draw at least 60 frames per second in many modern games, which previously could only be achieved in single titles.

Here, Nvidia also has not been without innovations. In the RTX video cards, thanks to the same tensor cores, the new DLSS technology or Deep Learning Super Sampling will be used (supersampling using deep AI training without specialized algorithms for individual tasks), which formally represents advanced smoothing, although in reality this is not quite like that.

Using neural networks, devices will improve the image, eliminating the need to include any other types of anti-aliasing – even the best in the performance-quality ratio, for example, TAA (fairly fast temporal smoothing, which does not require too much resources).

However, judging by the presentation of the company, in reality, when DLSS is enabled, the picture will be rendered in initially lower resolution, and already with the help of neural networks, the video card will artificially increase it to full 4K. This approach is a bit like 4K chess technology, popular on consoles, but thanks to the work of neural networks and tensor cores, PC players will be able to achieve a better image with fewer artifacts.

In addition, judging by the short rollers with a demonstration of the technology, the performance with the included DLSS will be much higher than with the standard 4K with anti-aliasing. For example, in the demo on the Unreal Engine with activated supersampling, the frame rate averaged 70 frames per second, whereas with TAA enabled – only 40. However, only off-screen recordings from the Gamescom exhibition are currently available, so to honestly compare DLSS operation with more familiar settings is still impossible.

Many PC users were skeptical about the capabilities of raytracing and did not understand why they even need to buy new expensive Nvidia video cards (especially in Russia, where the prices of many non-reference models with good cooling exceed100 thousand rubles). However, it should be taken into account that this is the first such breakthrough in the use of real-time tracing and it will take months or even years to complete its implementation.

The capabilities of the technology demonstrated by many developers are still farfrom ideal, and besides, many studios spent only a few weeks building the right builds, since they did not have access to the right devices. But now you can see the potential in the videos shown – some experts even believe that, unlike many past technologies (for example, Hairworks wool and Wool rendering or PhysX physics processing), in this case, monumental changes will take place in the graph, even if not right away

Of course, there are downsides. Probably now when using raytracing at best, it is worth waiting for performance at 60 fps at 1080p. This will certainly upset those who waited for emphasis on full-fledged 4K and a high frame rate in the new generation.

There is also a question with consoles, where in the current generation only Xbox One X can display 4K at 30 frames per second in many games, and 60 fps in general is only in a few projects. Future consoles Sony and Microsoft, as well as modern, according to rumors, will also work on AMD’s hybrid video chips, but the company is still silent about the potential support of raytracing in its future devices. But even if the support is still likely, the trace will mostly remain an “exclusive” of powerful PCs.

So far, it can only be noted that Nvidia is set up seriously – initially the company reported on joint work on a hybrid trace with Microsoft. Together, corporations are engaged in creating a special set of tools for developers DXR API for DirectX 12, which is responsible for graphics in Windows 10. Representatives of Epic Games also announced plans to add support for raytracing into their Unreal Engine, which will also contribute to the spread of technology and its use by developers.

Of course, in the near future the revolution should not be expected. Tracing support will already appear in the upcoming games this fall, but only in the base view. The main boom is probably to be expected only by mid-end of 2019, or even later, when projects will be released, which are now far from the final stages of development.

Nvidia is reluctant to share the details of the work of new technologies, so it remains only to wait for the release of video cards and the appearance of the results of the first tests in games, in order to understand whether they are really more powerful than the previous ones and that raytracing is capable of in real conditions.

On sale RTX 2080 and RTX 2080 Ti will appear on September 20, and the release date of the younger RTX 2070 is still unknown.

Tags

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button
Close
Close