RDNA4/gfx1201: HIP/ROCm Q2_0 ~13× slower than Vulkan (reproduced on ROCm 7.2.1 & 7.2.4)

[The-Monk]

Summary

On RDNA4 / gfx1201 (Radeon AI PRO R9700), the HIP/ROCm backend runs Ternary-Bonsai-8B Q2_0 prefill ~13× slower than the Vulkan backend built from the same tree (prism @ c85e97a). Vulkan is excellent; HIP is not. Full diagnosis below — it reproduces on ROCm 7.2.1 and 7.2.4 (latest stable), and even a scalar Vulkan path (matrix cores disabled) beats ROCm's best by ~4×, which points past MMQ into the gfx1201 rocBLAS/Tensile stack.

Environment

• GPU: 2× AMD Radeon AI PRO R9700 (RADV GFX1201), Vulkan 1.4.318
• llama.cpp: PrismML-Eng/llama.cpp prism branch @ c85e97a (PR vulkan: Q2_0 #32 merged)
• Vulkan build: -DGGML_VULKAN=ON · HIP build: -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1201
• ROCm 7.2.1 and 7.2.4 (both tested); rocBLAS 5.2.0 in both
• Model: prism-ml/Ternary-Bonsai-8B-gguf:Ternary-Bonsai-8B-Q2_0.gguf

Benchmark (llama-bench -ngl 99 -p 512 -n 128)

Backend / ROCm	pp512 (t/s)	tg128 (t/s)
Vulkan (KHR cooperative-matrix)	2775	132
Vulkan, all coopmat disabled (scalar fp16)	989	—
ROCm 7.2.1 (HIP)	250	99.5
ROCm 7.2.4 (HIP, latest stable)	249	89

Diagnosis (ruled out the usual suspects)

• Not a missing macro: amdclang++ --cuda-device-only --offload-arch=gfx1201 -dM -E defines __GFX12__ → RDNA4 + AMD_WMMA_AVAILABLE are ON.
• WMMA is dispatched: rocprofv3 shows void mul_mat_q<...> = 59.8% of GPU time (not a cuBLAS fallback).
• Not fallback selection: GGML_CUDA_FORCE_MMQ vs GGML_CUDA_FORCE_CUBLAS both ≈250.
• Not clocks: during compute the GPU sits at ~1340–1452 MHz / 48–76 W (cap 2350 MHz / 350 W); rocm-smi --setperflevel high doesn't change it (261 vs 250). The kernel stalls → low occupancy → DPM won't boost.
• Not a stale ROCm: reproduced on 7.2.4 (latest stable). rocBLAS is 5.2.0 in both 7.2.1 and 7.2.4, and the gfx1201 Tensile files are identical — notably the int8 GEMM is TensileLibrary..._fallback_gfx1201.hsaco (a generic fallback kernel, not perf-tuned).

The key signal

Vulkan on radv advertises only VK_KHR_cooperative_matrix and int dot: 0, so it does dequant Q2_0 → fp16 → KHR cooperative-matrix. HIP uses the integer MMQ (quantize→int8→WMMA) path. On gfx1201, even Vulkan's scalar fp16 path (989 t/s) is ~4× faster than ROCm's best (250) — and ROCm's own FORCE_CUBLAS (rocBLAS fp16 GEMM) is also 250. So this isn't purely a mul_mat_q tuning issue; the entire gfx1201 ROCm GEMM path (hand-written WMMA MMQ + vendor rocBLAS) is immature on this new silicon.

Reproduction

git clone -b prism https://github.com/PrismML-Eng/llama.cpp && cd llama.cpp
# Vulkan
cmake -B build -DGGML_VULKAN=ON && cmake --build build -j --target llama-bench
# HIP (gfx1201)
cmake -B build-hip -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1201 && cmake --build build-hip -j --target llama-bench
M=Ternary-Bonsai-8B-Q2_0.gguf
build/bin/llama-bench     -m $M -ngl 99 -p 512 -n 128   # Vulkan  ~2775 / 132
build-hip/bin/llama-bench -m $M -ngl 99 -p 512 -n 128   # ROCm    ~250  / 99

Question / ask

Is RDNA4 HIP/ROCm optimization for Q2_0 on the roadmap, or should Vulkan be the documented recommended backend for gfx1201? Happy to share the full rocprofv3 trace and test any patches. The scalar-Vulkan-beats-rocBLAS result suggests part of this belongs upstream in ROCm/rocBLAS (untuned gfx1201 Tensile) — I can file there too if useful.

(Report prepared with tooling assistance; all numbers measured on the hardware above.)

[The-Monk]

Correction — root cause is the fork's stale base, not ROCm

I built mainline ggml-org/llama.cpp HEAD (0eca4d4, 2026-06-30) for HIP (gfx1201) and re-ran a standard Qwen3-8B Q8_0 on the same hardware/ROCm. The result overturns my original diagnosis above:

Qwen3-8B Q8_0, gfx1201	pp512 (t/s)	tg (t/s)
Mainline HIP	4019	56
Mainline Vulkan	2865	58
This fork (prism) HIP	284	49
This fork (prism) Vulkan	2730	57

So:

• RDNA4/gfx1201 ROCm is not immature — mainline HIP is excellent (~4000 t/s prefill, faster than Vulkan). My earlier "ROCm 7.2.x/Tensile immaturity" conclusion was wrong.
• This fork's HIP path is ~14× slower than mainline. The fork's merge-base with mainline is >200 commits old (beyond a --depth 200 fetch), so its base predates mainline's RDNA4 HIP MMQ optimizations. The fork's Vulkan path is current (2730 ≈ mainline 2865); only HIP is stale.

Recommendation: rebase the prism branch (and the Q2_0 additions) onto current mainline — the HIP path would then inherit the ~14× speedup for free, and Q2_0 on ROCm would run at mainline-class prefill instead of ~284 t/s. Happy to help test a rebased branch.

Apologies for the incorrect root-cause in the original report — measuring against actual mainline HEAD is what surfaced it.

[The-Monk]

FYI — I implemented Q2_0 against current mainline ggml-org/llama.cpp (the fast HIP path from #49): CPU + CUDA/HIP dequant, MMVQ, and MMQ. Correctness-verified, and on gfx1201 it runs pp512 3877 / tg128 114 t/s (vs 250/99 on this fork). Draft PR: ggml-org#25188 — branch The-Monk/llama.cpp:cuda-q2_0-ternary. Feel free to use it for the upstream PR you mentioned.