The [1]AMD GFX1250 IP is what's used by the new Instinct MI450 series accelerators. That's been around within the Linux kernel, AMDGPU LLVM compiler back-end, and related codebases the past year. Interestingly merged yesterday is a new "gfx1250-strict" target.

Opened up and already merged yesterday to the LLVM compiler codebase is the new "gfx1250-strict" GPU target. Going through the initial addition patch, "GFX1250_STRICT" enables all of the same ISA capabilities and the like as the vanilla "GFX1250" target and copying other existing GFX1250 definition. Presumably the new "strict" functionality will be built out over future, yet-to-be-published patches. There is also only this "strict" or "12.5.0s" target and not any similar strict definition for the GFX1251 IP version.

The only LLVM pull request thus far is [2]this one that defines the gfx1250-strict target. As always though, I am constantly on the lookout for new and interesting open-source code.

This is the first time I am seeing any AMD GPU "strict" IP target in my monitoring.

With the lack of information thus far from the basic enablement patches and no pull request description, it's not clear if this new "gfx1250-strict" is ultimately for a new hardware product like a safety-critical AMD accelerator operating in a more strict/hardened environment or if it's for an alternative compiler target for the existing Instinct MI450 series hardware when running workloads that are more regulated or requiring strict conformance. We'll see how future patches build out this target.



[1] https://www.phoronix.com/search/GFX1250

[2] https://github.com/llvm/llvm-project/pull/218543