An AMD GPU Architecture Patent has been discovered on the internet and it shows some features that you can expect to see post AMD Navi launch. Before getting into this keep in mind that companies often get patents on technologies that never come out. There is a possibility that this technology will never see the light of day but one can hope.
While the patent itself is pretty complicated the following is the breakdown that has been provided by Dijky.
Each ALU gets a “sister” ALU next to it.
Both ALUs are controlled by a VLIW2 instruction (i.e. one big instruction that contains two individual instructions, one for each ALU).
Both ALUs’ outputs go into a Destination Operand Cache (Do$). From there, the outputs can go back into the VGPRs (vector register file), or they can be directly forwarded back into the two ALUs, saving read and write operations on the VGPRs.
There are also two versions of how an overall “Compute Unit” is assembled from either two or four of these Super-SIMD blocks and other supporting hardware (AFAIK similar to GCN CUs, except there is now also a “compact” version)
The ALUs in one Super-SIMD can be a combination of (vectorized) “full ALUs”, “core ALUs” (presumably only implementing a few important instructions), and “transcendental ALUs” (takes longer, for stuff like sin, cos, sqrt, exp etc.).
There is also vectorized “side ALU” that can aid both ALUs of a Super-SIMD to perform non-essential operations.
Previously, AMD/ATi has used VLIW architectures with 5 or 4 instructions, since the introduction of unified shaders. Before that, GPUs had fixed-function pipelines. GCN was quite a radical shift in terms of using RISC instructions instead.
This patent seems to be a sort of compromise back to a combination of GCN and VLIW (now with 2 instructions), and potentially some architecture specialization at the ALU and CU level, This is speculation, AMD could just try to patent a broad range of designs.
It is interesting to note that similar features have been available on the Nvidia side of things since Fermi and Maxwell. You can check out the patent here in order to learn more. This could be a legal thing but all we can do is wait and see. For more information regarding the AMD GPU Architecture Patent stay tuned to SegmentNext.
Let us know what you think about this AMD GPU Architecture Patent and whether or not you think these features will ever see the light of day.