AMD GFX1310, aka RDNA5 appears in LLVM update
AMD may be making it easier for RDNA 5 GPUs to use a feature called Dual Issue. Dual Issue lets the GPU do two math operations at the same time in some cases, instead of just one. A recent LLVM code change for gfx13 , which is tied to AMD’s upcoming gfx1310 architecture, suggests RDNA 5 could use that feature more often than older designs. The change appeared in an LLVM pull request, and Coelacanth’s Dream was the first to explain why it matters.
The main point is that AMD’s earlier Dual Issue setup had a lot of limits. It only worked with a smaller list of instructions, needed Wave32 mode, and had extra register rules that could block it from working. One important FP32 instruction, V_FMA_F32, was not part of the older path, so some benchmark tools such as vkpeak could show only about half of the GPU’s claimed peak FP32 performance in certain tests.
RDNA 5 (gfx1310)VOPD added instructions such as V_{MAX,MIN}_I32_e32, V_SUB_U32_e32, V_LSHRREV_B32_e32, V_ASHRREV_I32_e32extracting the maximum and minimum values of integers, subtraction instructions, and bit shift instructions.
— Coelacanth’s Dream
The new LLVM code shows AMD is adding more support for gfx13 through new entries called GFX13GenD and GFX13GenD3 . That means RDNA 5 is getting updated Dual Issue instruction handling, including support for more operations than before. The patch also adds gfx13-specific test files, so this looks like an actual feature update, not just an early placeholder.
What this means for gamers (an devs) is that RDNA 5 may be able to keep Dual Issue active in more situations, including some workloads that older RDNA designs could not handle as well. That would help certain FP32 benchmarks get closer to the GPU’s full advertised throughput, even if AMD does not raise the peak FP32 rate itself. In other words, the GPU may get better at reaching its existing limit, not necessarily pushing that limit higher.
Of course, that is only a theory; as always, implementations will determine how this works in reality.
Source: Coelacanth’s Dream
