ollama - 💡(How to fix) Fix ROCm performance regression on gfx1201 (RX 9070) between 0.24.0 and 0.30.6 on Windows 10

What is the issue?

Significant token generation performance regression on AMD RX 9070 (gfx1201) between Ollama 0.24.0 and 0.30.6 on Windows 10. On 0.24.0 with HIP SDK 6.4.2 and ROCBLAS_TENSILE_LIBPATH pointing to the external library: ~53 tok/s on qwen3:14b (Q4). On 0.30.6 with native bundled ROCm (rocm_v7_1): ~42 tok/s on the same model and hardware. Both versions confirm GPU detection via server.log: library=ROCm compute=gfx1201 description="AMD Radeon RX 9070". This is not a CPU fallback issue; the regression is purely in inference throughput on confirmed GPU compute.

Note: on 0.24.0, the ROCBLAS_TENSILE_LIBPATH workaround pointing to HIP SDK 6.4.2 libraries was required to enable ROCm detection at all for gfx1201. This workaround is no longer needed on 0.30.6 which detects the GPU natively, however inference performance has regressed significantly compared to 0.24.0 with the external library.

Environment: OS: Windows 10 GPU: AMD Radeon RX 9070 (gfx1201) CPU: AMD Ryzen 5600 Model tested: qwen3:14b Ollama 0.24.0: ~53 tok/s Ollama 0.30.6: ~42 tok/s

Relevant log output

time=2026-06-08T13:06:36.722+02:00 level=INFO source=types.go:32 msg="inference compute" id=0 filter_id=0 library=ROCm compute=gfx1201 name=ROCm0 description="AMD Radeon RX 9070" libdirs=ollama,rocm_v7_1 driver=0.0 pci_id=0000:10:00.0 type=discrete total="15.9 GiB" available="15.8 GiB"

time=2026-06-08T13:38:47.288+02:00 level=INFO source=types.go:42 msg="inference compute" id=0 filter_id=0 library=ROCm
compute=gfx1201 name=ROCm0 description="AMD Radeon RX 9070" libdirs=ollama,rocm driver=60450.10 pci_id=0000:10:00.0
type=discrete total="15.9 GiB" available="13.1 GiB"

OS

Windows 10

GPU

AMD Radeon RX 9070 (gfx1201)

CPU

AMD Ryzen 5600

Ollama version

Ollama 0.30.6