AMD has been using the term Compute Unit for over a decade, which encompasses a full GPU core or what we call Shader Unit. However, future challenges to achieve more performance mean that for future architectures internal changes will have to be made. Which in the case of AMD is surprising, especially since for generations they have kept the same type of computing units.
Now, and against all odds, it looks like Navi 31 is going to have twice as many ALUs in FP32 compared to Navi 21, which means reducing the difference on this aspect compared to NVIDIA.
The final farewell to compute units in RDNA 3?
In Beyond3D, in the thread dedicated to RDNA 3, there is an interesting conversation, in which a certain Bondrewd assures that there are no more “Compute Units” but that the minimum unit will be the WGP. Which comes in response to the latest information on RDNA 3.
To understand this claim, we must consider how in RDNA and RDNA 2, each workgroup is made up of two computing units, where the two units share common elements such as the scheduler, first-level instruction cache, and local shared memory. On the other hand, other elements such as the texture filtering units, the Load / Store units and the SIMD units of each Computing Unit continue to be maintained.
With andl Dual threads in FP32 for RDNA 3 As leaked a few days ago, the question revolves around what changes AMD can make to achieve this. We don’t think AMD will rule out the classic finite compute unit, but we will see a difference in the ratio between the ALUs in FP32 and the texture units, where it will have the same as that of the NVIDIA RTX 30, that is. that is, 32 per ALU in FP32 for each texture unit for a total of 128 ALUs per Compute Units and 256 per Workgroup or WGP.
A design copied from NVIDIA or a different concept?
NVIDIA and AMD pose SIMD units differently from their shader units. In the case of NVIDIA, we have two sets of different SIMD units in each sub-core of its SM, the first of them with 16 ALUs in FP32 and another collection of 16 ALUs in FP32 or Int32, where the two ALU types are switched with each other. .
In exchange AMD takes a different approach, the diagrams do not show the Int32 units, but they are and are all switched with an Int32 unit. The two approaches are therefore different, but what they mark is an increase in the rate of computing power. Which makes sense with the adoption of Ray Tracing and the fact that it requires the use of brute force computation and a lot less fixed pixel fill and texturing units.
For what AMD for its RX 7000 would have raised an entirely different computing unit, with twice the computing power of the RX 6000. It is said that the first chip will be the monolithic Navi 33, if the rumors are true, we would talk about a potential GPU of 40 computing units but with a computing capacity similar to that of the RX 6900 XT.