Computer

NVIDIA will leave graphics cards for gaming and it’s no joke

Real-time 3D games today are 99% of games, but on computers they arrived thanks to the popular Voodoo Graphics in its day and the fact of having a chip that was in charge of a common function. We are talking about the raster unit, which is found today in all graphics chips, but which will disappear over time. What consequences will this have on the hardware of the future?

GPUs are changing

GPUs are specialized graphics processors, with some ability to work in general purpose. This is because their cores have long since ceased to be for specialized tasks to be used for general tasks. Hence, many of the world’s great supercomputers use them today for large-scale scientific and military simulations, as well as other booming applications such as artificial intelligence.

However, there are still a series of units that carry out specific tasks, but that are necessary to be able to generate the graphics. Their job is not only to free the main nuclei from them, but to work in parallel. Its biggest advantage is that by having a fixed or specific task to do, they require fewer transistors to build and, therefore, are less costly and consume less than making the main core take care of the task.

However, it can happen that a specialized piece of hardware becomes outdated, either because it doesn’t keep up with the performance advances of other items or because it works in a less than ideal way. Well, we are finding that in the main chip of the graphics cards a specific type of unit is appearing and another type is disappearing.

NVIDIA Top-of-the-range graphics Without fans or heatsink

Farewell to the raster unit

If you look at the specifications of any graphics card, you will see that one that is less and less important is the one that talks about “triangles per second”. Many people think that it is the amount of these that are put on the screen and others will believe that they are the amount that is calculated. Both are false due to the fact that this depends on the application that we are running. Moreover, it is a fixed rate and if we notice it is always the clock speed for the number of raster units. At least today.

These units were advancing in performance, from requiring several clock cycles to rasterize a triangle, to doing it at the current speed, but they have not evolved for ten years. His work? Project the 3D world calculated in vertices to a 2D surface made up of pixels, the screen, to which the color will later be calculated. Therefore, it is an essential unit, since when we make a 3D scene, each and every one of the triangles ends up becoming pixels; however, said unit could soon say goodbye.

The reason? Your limitations

The problem with the raster units is that they are not designed to work with very small triangles, that is, those that are rasterized in a few pixels. Furthermore, when an object is made up of only a few pixels, what the rasterizer does is conclude that it is too far away and mark it for deletion. Of course, before checking its distance to the camera with the depth buffer to eliminate it from the scene and that it does not have to be calculated. This is also done with objects behind a larger one, which causes problems when the frontmost object is transparent to some degree. Which causes problems, but that’s another matter entirely.

Rasterizer Efficiency

The problem? We are moving into a world where geometry is being used at extreme levels for detailing characters, objects, and settings. Which means triangle rates that current raster units cannot support. Which is a bottleneck, but the problem is that they can’t work well with small triangles. In the sense that if we have 100 triangles of 50 pixels, this does not translate into 200 of 25 pixels. Therefore, its efficiency decreases as the polygons with which it works are smaller.

What is the solution?

So much so, that the people at Epic Games in the face of the creation of Unreal Engine 5 have had to create raster units using computer shaders for it. That is, the GPU cores doing the job of a specialized function unit much better. Which endangers its future. At the moment it has not been ruled out, however, it already has the sword of Damocles on it, as well as the tessellation or surface subdivision unit.

Detail Geometry Engine UE5In an interview conducted in May 2020 with Brian Karis, the graphics programmer stated that they have developed two kinds of software triangle rasterizers for Unreal Engine 5. This means that Epic’s new engine, which will be used by dozens of games in the industry, already has the ability to dispense with the raster unit. That is, replace one of the GPU cores to replace each of these units and gain more performance with it.

The vast majority of triangles are rasterized by software using highly specialized computer shaders, which have been designed to take advantage of what we can benefit from. As a result, we’ve been able to leave hardware rasterizers in the dust on this specific task. Software rasterization is a major element of Nanite that allows us to do what it does. We can’t beat hardware rasterizers in every case, so we’ll use them when we determine it’s the fastest path.

White and in the bottle, as you can read. They will disappear due to the fact that they are transistors that NVIDIA, Intel and AMD can take advantage of for other things that will become more important in the future.

Why will the organization of graphics cards change?

If you look at the diagrams of any GPU you will see that the cores are arranged in blocks around a rasterizer. This is because they send both data to these units and also receive them, depending on which stage of the 3D pipeline we are talking about. So removing it is a reorganization. Currently, the ideal size for 100% efficiency of raster units is 48 pixels for NVIDIA and 64 for AMD. This also limits the number of cores in models designed for gaming. Which, in the case of the brand in green, can be seen by comparing its chips for high performance computing and gaming.

H100 AD102 CPG compared

As can be seen, the H100 GPC does not have a Raster Engine, which allows it to not have such a fixed organization and, therefore, limited. This change will allow NVIDIA, and also AMD, not to have to design two different designs for HPC and Gaming, but instead will be able to pull from a universal base model in terms of design. From which you can iterate. Today, regardless of whether we are talking about a central processor or a graphics processor, we find that the intercommunication is between 2/3 and 3/4 of the total chip and the fact of having to work on two different chips is daunting.

Let’s not forget that each new node is more transistors, this is more parts and more engineers to hire. It will reach the point where the same graphics cards that are offered for scientific work will be the high-end for gaming, due to the fact that it will not be profitable to make two different models and the elimination of the raster unit will be key in this entire development process. unification.

Related Articles

Leave a Reply

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