Recently, information on domestic GPUs, Intel-only GPUs, and next-generation GPUs such as Nvidia RTX 40 series and AMD RX 7000 has been continuously exposed in the media. But these GPUs have not been officially released, certainly can not use 3DMark and various games actually run, running score to show the ability, related reports in the "comparable to xxxx", "stronger than xxxxxxxx" in the relevant reports, how is it from and where does it come from?
The current GPU is composed of many small processing cores, or stream processors, which are only responsible for processing one floating-point data per clock cycle, so the total number of floating-point operations is the number of cores × clock cycles (the current common GPU floating-point operation unit is generally TFLOPS, that is, how many trillions of floating-point operations per second. ）。 And because the core can now process one double-precision floating-point data at a time, which is equivalent to two single-precision floating-point data, it ×2 to get the number of floating-point operations of the GPU.
At present, the performance comparison between domestic GPUs, NVIDIA RTX 40 series and AMD RX 7000 series and current GPUs and graphics cards is actually the theoretical floating-point computing performance calculated according to the frequency and number of stream processor configurations revealed. Because the current image is divided into pixels to process, each point of color must be floating-point operation, and then combined into a picture, so this floating-point computing power can indeed represent the graphics card or GPU image processing power.
Because the more pixels are processed per second, the more screens can be processed per second in the same image quality and the same resolution, and the higher the frame rate (number of screens per second) of the game. For graphics cards that use the same generation, especially the same core, calculate its floating-point computing power, and basically understand the generation speed of the game screen. Of course, this is the case that accessories such as CPU, memory, output interface, and display are not lagging behind.
But for different generations or even different architectures of GPUs, this comparison is not appropriate, such as RTX 3080 has twice the number of floating-point operations of the RTX 2080 Ti, can the frame rate reach twice that of the RTX 2080 Ti? This involves another problem, that is, the efficiency of the core, because no one can guarantee that all the core or stream processor can always be fully loaded and run efficiently, and its actual play also takes into account the allocation of the front end, the synthesis of the back end, the cooperation of the video memory data and other units.
The RTX 30 series is even more "problematic" by allowing each of the core units of integer operations to participate in floating-point operations, resulting in exponentially doubling the "theoretical" computing power. However, in non-"professional" work, the floating-point operation efficiency of integer units is definitely not as good as that of professional floating-point units, so the efficiency is greatly reduced. The resulting doubling of theoretical floating-point performance results in less than a 40% actual frame rate increase. So how the RTX 40 and RX 7000 are, in addition to the floating-point computing performance, it also depends on whether there are major changes in architecture and efficiency.
In fact, there is a difference between the theoretical performance of domestic GPUs and Intel GPUs not mentioned in the previous comparison, that is, they have begun to be intensively tested, so in addition to floating-point computing, there is also an OpenGL computing power. The relevant tests can reflect the efficiency of the GPU architecture to a certain extent, but they cannot be completely communicated with the game performance, because the game requires consumer-grade 3D technologies such as Direct 3D, which is not necessarily the same as the operation of OpenGL, and can only be said to have a greater reference significance than floating-point computing.