AMD has been using the term Compute Unit, which encompasses an entire GPU core or what we call a Shader Unit, for over a decade. However, the challenges of the future to get more performance means that for future architectures, internal changes will have to be made. Which in AMD’s case is surprising, especially if we take into account that for generations they kept the same type of Compute Units.
Now, and against all odds, it seems that Navi 31 is going to have twice as many ALUs in FP32 compared to Navi 21, which means cutting the difference in this aspect compared to NVIDIA.
The definitive goodbye to Compute Units in RDNA 3?
In Beyond3D, in the thread dedicated to RDNA 3 there is an interesting conversation, in it a guy named Bondrewd assures that there are no more “Compute Units” but the minimum unit will be the WGP. This comes as a response to the latest information about RDNA 3.
To understand this statement we have to take into account that in RDNA and RDNA 2 each Workgroup is composed by two Compute Units, where both units share common elements such as the scheduler, the first level instruction cache and the shared local memory. On the other hand, other elements such as the texture filtering units, the Load/Store units and the SIMD units of each Compute Unit remain the same.
With twice as many execution units in FP32 for RDNA 3 as leaked a few days ago, the question is about what changes AMD can make to achieve this. We don’t think AMD will ditch the classic finished Compute Unit, but we are going to see a difference in the ratio of FP32 ALUs to texture units, where it will be the same as the NVIDIA RTX 30, i.e. 32 per FP32 ALUs per 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 approach the SIMD units of their shader units differently. In NVIDIA’s case we have two different sets of SIMD units in each subcore of their SM, the first one with 16 ALUs in FP32 and another collection of 16 ALUs in FP32 or Int32, where the two types of ALUs are switched with each other.
On the other hand AMD has a different approach, in the diagrams you can’t see the Int32 units, but they are and they are all switched. with an Int32 unit. So the two approaches are different, but what they do mark is an increase in the rate of power per computation. Which makes sense with the adoption of Ray Tracing and the fact that this requires the use of brute force per compute and much less fixed units for texturing and pixel filling.
So AMD for its RX 7000 would have proposed a completely different Compute Unit, with twice the computational 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 be talking about a potential GPU with 40 Compute Units but with a computing power similar to that of the RX 6900 XT.