[Feature]: [gfx110x / gfx115x / gfx120x] Prioritize full Composable Kernel (CK) enablement for MIOpen — revive closed PR #2755 #9838

Suggestion Description

Summary

Please prioritize (and revive) full Composable Kernel (CK) support for MIOpen on the gfx110x, gfx115x, and gfx120x families.

PR #2755 (“Enable CK for MIOpen support on gfx110x, gfx120x, and gfx115x”) was closed due to inactivity and an extremely outdated branch. This is a high-impact gap for a large and growing user base on RDNA 3, RDNA 3.5, and RDNA 4 hardware.

Related work already exists in rocm-libraries (notably around gfx120X in PR #5462, which demonstrated substantial measured gains) and should be coordinated with TheRock so that nightlies and releases actually ship CK-enabled MIOpen for these architectures.

Background

• Closed PR: Enable CK for MIOpen support on gfx110x, gfx120x, and gfx115x #2755
  (opened January 2026, closed November 2026 for inactivity / outdated branch)
• Strong related work in rocm-libraries:
  feat(miopen): Add more GEMM convolution solvers from CK for gfx1201 (RDNA4) rocm-libraries#5462
  (focus on enabling gfx120X static-CK implicit GEMM convolution solvers)

Measured gains on gfx1201 (from that work) include:

• ResNet18 FP16: +111% throughput
• ResNet50 FP16: +52% throughput

Without the CK paths, these architectures fall back to much slower generic solvers (and can hit issues such as FP32 exhaustive search aborts).

gfx110x, gfx115x, and gfx120x are already present in SUPPORTED_GPUS.md with varying levels of “Build Passing / Sanity Tested / Release Ready” status on Linux and Windows, yet the high-performance CK convolution paths that make MIOpen competitive are not fully enabled and stabilized in TheRock distributions.

Why this matters for users

These three families cover a very wide range of hardware used for AI and compute workloads:

• gfx110x (RDNA 3) — RX 7000 series and related cards
• gfx115x (RDNA 3.5) — Strix Point / Strix Halo APUs and related
• gfx120x (RDNA 4) — RX 9000 series and AI-oriented workstation cards (including 32 GB models)

Enabling mature CK-backed MIOpen solvers unlocks significantly better training and convolution-heavy inference performance. This is especially valuable for local AI, fine-tuning, diffusion models, and multi-GPU setups on accessible hardware. Users currently leave performance on the table on cards that are otherwise well-supported by ROCm.

Why this is valuable for the ROCm ecosystem

1. Much broader addressable market — RDNA 3 / 3.5 / 4 consumer, APU, and workstation cards are far more accessible than Instinct. Strong performance here multiplies the number of developers and researchers who can use, test, and contribute to ROCm.
2. Competitive positioning — Competitive convolution and training performance on these widely available GPUs is a practical on-ramp to the full ROCm stack and a clear differentiator.
3. Reduced fragmentation — Coordinating TheRock + rocm-libraries CK enablement avoids partial support that exists upstream but is not reliably available in nightlies/releases.
4. Linux + Windows parity — Both platforms already have substantial coverage for these families. Completing CK/MIOpen support strengthens the overall story.
5. Contributor magnet — Users with RDNA 3/3.5/4 hardware are highly motivated to test, report, and contribute. Clear prioritization attracts more external help.

Proposed next steps

• Rebase / revive the intent of Enable CK for MIOpen support on gfx110x, gfx120x, and gfx115x #2755 (or open a fresh PR) so that TheRock actually builds and ships MIOpen with CK enabled for gfx110x, gfx115x, and gfx120x.
• Align with the changes in rocm-libraries#5462 (and any related follow-ups) so the high-performance static-CK solvers are available.
• Ensure nightlies for the relevant families (gfx110X-all, gfx115X, gfx120X-all, etc.) contain a CK-enabled MIOpen (verify via MIOPEN_USE_COMPOSABLEKERNEL and solver listing).
• Expand CI coverage / sanity tests on real hardware for these architectures where feasible.
• Document expected performance characteristics for common AI workloads once the path is stable.

Offer of help

I have RDNA 4 workstation hardware and can provide:

• Real-hardware testing (Linux, and Windows if useful)
• Performance numbers on representative training and inference workloads
• Bug reports and validation of nightlies once the feature lands

Happy to help test any rebased branch or new PR. Community members with gfx110x and gfx115x hardware can similarly contribute validation.

Thank you for the continued work on these architectures — making the high-performance CK paths first-class would be a major win for both the ROCm ecosystem and the large community of Radeon / Ryzen AI users.

Operating System

No response

GPU

No response

ROCm Component

MIOpen / Composable Kernel (CK)