One of the biggest problems that exist when it comes to gaming performance on PC has to do with the fact that the GPUs inside different graphics cards of all brands do not speak the same language and the use of a real-time translator is necessary. , which prevents you from taking full advantage of them. Why don’t they all speak the same language if they are going to run the same games and programs, at least in the visual part?
Why is a Universal GPU ISA necessary?
When it comes to executing programs, it is normal for all CPUs compatible with an ISA to be able to execute the programs designed for it without problems, unless we are talking about extensions to the set of registers and instructions thought long after the launch of said program. That is, the binaries work perfectly and it is not necessary to compile them again for each new chip and architecture that appears on the market.
Unfortunately, this is not the case with GPUs, where Intel, AMD, and NVIDIA not only use a different set of registers and instructions not only between themselves, but also between generations of graphics cards. This means that while the game is running, the different shader programs have to be compiled to save them in a file on the hard drive that will be recovered later. This causes the first games when reaching certain areas of a game for the first time to have huge frame drops on PC, although that is not the only problem
The fact of having to store so much code written in high-level shaders, whether in HLSL, GLSL or any other language like that, requires a larger amount of video memory. All this makes us wonder why both NVIDIA and AMD, and if we also have the rest of GPU manufacturers and designers, do not want to make a universal GPU ISA, which will allow shaders to be compiled on any platform for better performance, avoiding the shader compilation issues that exist on PC.
The Steam Deck is the perfect example
Valve’s console has a series of games with a profile designed for it that, among other things, includes shaders already compiled for the GPU ISA it uses, that is, RDNA 2, which, curiously, is the same as in consoles such as Xbox series and Play Station 5. So there is no shader pre-compile process and all its associated problems. However, on PC that is impossible, given the huge number of graphics card models that exist and the differences between them.
Why hasn’t this solution been applied?
One of the historical curiosities today occurred with the first Xbox, where the GPU vendor was envious and took advantage of its advantage to price Microsoft well above, but could afford what led Redmond to seek reverse engineering or rather use a chip from its rival ATI for its console, but given the exclusivity of the ISA of each of the chips, that was not carried out, although it helped ATI Technologies to become the supplier of the Xbox graphics chip 360.
Currently, ATI Technologies is part of AMD, it is the current Radeon Technology Group, and of the three big brands, only Lisa Su’s has the capacity to provide a processor and graphics card in a single chip with the power necessary for a console. NVIDIA completely lacks an x86 license and Intel is nobody at the moment when it comes to graphics cards. So a universal GPU ISA wouldn’t be a problem, and consoles aren’t the point.
Rather it is the ecosystem that NVIDIA has created through CUDA today and with which it has absolute dominance in various markets. If a universal ISA was made for GPUs, then the green mark would lose its advantage because applications and libraries created out of envy for its hardware would run flawlessly on the competition, although it would have temporary first-mover advantage.
