softab

Whisper.cpp Benchmarks on AMD Strix Halo (gfx1151)

Last Updated: 2026-02-02 Hardware: AMD Ryzen AI Max+ 395, Radeon 8060S (40 CU, gfx1151), 128GB unified memory Test Audio: 11s JFK “Ask not” speech (test-audio.wav) Model: ggml-base.en.bin (74M parameters)

Executive Summary

• Best config: ROCm 7.2 HIP with SDMA=1 (~447ms) or hipBLASLt=0 (~459ms)
• Vulkan RADV/AMDVLK both work (~533-548ms) - slightly slower than HIP
• ROCm 6.2/6.4.4 hang after model load - avoid these versions
• SDMA=1 slightly faster than SDMA=0 for whisper (opposite of PyTorch!)

Benchmark Results (11s audio → transcription time)

Image	Backend	Total Time	Realtime	Status
whisper-ablation-sdma-1	ROCm 7.2 HIP	447 ms	24.6x	✅ PASS
whisper-hipblaslt-0	ROCm 7.2 HIP	459 ms	24.0x	✅ PASS
whisper-ablation-sdma-0	ROCm 7.2 HIP	461 ms	23.9x	✅ PASS
whisper-hip-rocm72-gfx1151	ROCm 7.2 HIP	487 ms	22.6x	✅ PASS
whisper-vulkan-amdvlk	Vulkan AMDVLK	533 ms	20.6x	✅ PASS
whisper-vulkan-radv	Vulkan RADV	547 ms	20.1x	✅ PASS
softab-toolbox:whisper-vulkan-radv	Vulkan RADV	548 ms	20.1x	✅ PASS
whisper-hip-rocm62-gfx1151	ROCm 6.2 HIP	-	-	❌ HANG
whisper-hip-rocm644-gfx1151	ROCm 6.4.4 HIP	-	-	❌ HANG
whisper-therock-gfx1151	TheRock	-	-	❌ NO CLI
whisper-hipblaslt-1	ROCm HIP	-	-	❌ NO CLI

Realtime = audio_duration / transcription_time (higher = faster)

Key Findings

1. ROCm 7.2 HIP is Fastest

• ~450-490ms vs ~530-550ms for Vulkan
• ~15% faster than Vulkan backends

2. SDMA Behavior Differs from PyTorch

| Setting | PyTorch | Whisper | |———|———|———| | SDMA=0 | Recommended | Slightly slower | | SDMA=1 | May cause artifacts | Slightly faster |

For whisper, SDMA=1 appears safe and gives ~3% better performance.

3. hipBLASLt=0 Works Well

Unlike PyTorch where hipBLASLt=0 gives +20%, whisper shows minimal difference.

4. ROCm 6.x Versions Hang

Both ROCm 6.2 and 6.4.4 hang after loading the model. Use ROCm 7.2+.

5. Vulkan: AMDVLK Slightly Faster

• AMDVLK: 533ms
• RADV: 547ms
• Opposite of llama.cpp where RADV is usually better

Image Compatibility

Status	Images
✅ Working	whisper-hip-rocm72-, whisper-ablation-, whisper-hipblaslt-0, whisper-vulkan-*
❌ Hang	whisper-hip-rocm62-, whisper-hip-rocm644-
❌ Missing CLI	whisper-therock-*, whisper-hipblaslt-1

Recommended Images

For production (stability):

softab:whisper-vulkan-radv              # Vulkan, most stable
softab-toolbox:whisper-vulkan-radv      # Toolbox variant

For performance (fastest):

softab:whisper-ablation-sdma-1          # ROCm 7.2, SDMA=1, fastest
softab:whisper-hipblaslt-0              # ROCm 7.2, hipBLASLt disabled

Running Benchmarks

# Quick test (11s audio)
./benchmarks/whisper-simple.sh softab:whisper-vulkan-radv /data/models/test-audio.wav

# With timing output
podman run --rm \
  --device=/dev/kfd --device=/dev/dri \
  --ipc=host \
  --security-opt seccomp=unconfined \
  --security-opt label=disable \
  -v /data/models:/models:ro \
  softab:whisper-hip-rocm72-gfx1151 \
  whisper-cli \
    -m /models/ggml-base.en.bin \
    -f /models/test-audio.wav \
    --no-timestamps

Comparison: Whisper vs PyTorch Recommendations

Setting	PyTorch	Whisper	Notes
ROCBLAS_USE_HIPBLASLT	0 (disable)	0 or 1	Critical for PyTorch, minimal for Whisper
HSA_ENABLE_SDMA	0 (disable)	1 (enable)	Opposite recommendations!
ROCm version	TheRock 7.11	ROCm 7.2	Both work, whisper less picky
Vulkan driver	RADV	AMDVLK slightly faster	Different optimal drivers

Known Issues

1. ROCm 6.x hangs - Model loads but inference never starts
2. Some images missing whisper-cli - Build issue, use alternative image
3. TheRock images use different binary name - May be main instead of whisper-cli

---

See also:

• Ablations Added - Planned ablation tests
• Benchmark Scripts - Test scripts