NVIDIA's stream processors each have a complete ALU (which can be understood as mathematics, logic and other operations) function. NVIDIA has adopted full scalar design since G80, and all operations are converted into scalar calculations. However, once the 4D vector operation is encountered, it takes four operations to complete, so the shader frequency of NVIDIA graphics card is almost twice the core frequency, which just makes up for this shortcoming.
NV's stream processors have complete ALU function, so each stream processor consumes a lot of transistors, which is very expensive. Coupled with the current CUDA function, the number of transistors is much larger than AMD-ATI.
AMD-ATI starts with RV670, and the stream processor is a set of five fixed unified rendering units and a combination of 4D vector+1D scalar. Four of them can only perform MADD (multiply and add) operations, and 1 can perform "super operations (functions and other operations)". Because five spus are fixed as a group and cannot be split, when encountering pure scalar operation, four spus will be idle and other SP combination auxiliary operations cannot be added. But in other words, if each SPU is properly allocated and fully works, then the efficiency of AMD graphics cards can be very high. This is also the player's recognition that the performance of A-card driver is higher than that of N-card, but it is precisely for this reason that the design of A-card driver is very difficult, and it is also very difficult to optimize the game for A-card.
Let's talk about PhysX physical acceleration in NVIDIA. NVIDIA-PhysX physical acceleration is different from PhysX physical acceleration. It is better to have an independent floating-point processor for PhysX physical acceleration. Because GPU's floating-point operation is stronger than CPU's, NVIDIA chose to let GPU play the role of this floating-point processor. The full name of PhysX currently used by NVIDIA is PhysX for CUDA, which means that the PhysX currently used is based on CUDA. You must have a GPU supporting CUDA to use NVIDIA's PhysX physical acceleration, so that AMD-ATI can cool down. NVIDIA now wants to use PhysX for CUDA to suppress AMD, instead of turning PhysX physical acceleration into universal technology.
Physical acceleration and Havok have been acquired by Intel. Havok can be realized with only one CPU, and both INTEL and AMD's CPUs support Havok physical acceleration. Judging from the usage, the proportion of Havok is overwhelming. However, the purpose of Intel's acquisition of Havok is to share a piece of the GPU field, so AMD will not be cheap, and AMD can cool down again and again.
NVIDIA has PhysX, Intel has Havok, and AMD has nothing. But now that OPENCL has appeared, AMD seems to finally see a glimmer of light. Physical acceleration will be a fire in the future, but both PhysX and Havok are still lack of universality, and a standard will only hinder the development of the industry. However, there is an open OPENCL standard that is expected to become a new universal standard. Invista also intends to transfer CUDA to OPENCL. If so, AMD-ATI will also be able to use PhysX, but OPENCL seems to be unpopular with Microsoft, so who knows what will happen in the future? ...
As far as the current GPU performance is concerned, it is not enough for PhysX to use GPU acceleration, because the GPU is basically full during the 3D game, and it is impossible to extract enough performance to do PhysX operation. However, in 3D games, the CPU is rarely fully loaded for a long time, which is one of the advantages of Havok CPU acceleration.
Therefore, many people should not always regard PhysX as a noble standard. Now PhysX for CUDA is a short-lived species. If NVIDIA doesn't find another way, PhysX for CUDA will surely die. The physical acceleration of AMD-ATI is still blank, although it has the research and development capabilities of both CPU and GPU.