AMD R9700 owner: ROCm on Windows cripples local i2v, Vulkan runs fine
vladomkd · reddit · 2026-10-06
- A Reddit user reports that running i2v workflows (MiniMax H3, LTX 2.5) in ComfyUI on an AsRock AI Pro R9700 (32GB VRAM) under Windows 11 consumes 80-90GB of combined VRAM+RAM and takes 3-4 hours to generate a 20-second video.
- The GPU draws its full 300W but compute utilization stays near 0% — time is spent shuffling data between RAM and VRAM; the --gpu-only flag doesn't help.
- The same models run fine GPU-only via the Vulkan runtime, and ROCm shows similar degradation in LM Studio, leading the user to conclude ROCm on Windows is the problem. They ask how to force ComfyUI onto Vulkan and which quantized i2v models fit in 32GB VRAM.
More from Infra
- Strata v0.1.40 adds Strix Halo support, multi-GPU batching and decode improvements — lxfater · 2026-10-06
- BlackBerry's QNX hits record $80.3M revenue, up 27%, betting on AI-era safety platform — alysha_lobo · 2026-10-06
- Long-running benchmarks find Strata inference server failing full-build scenarios — julianharris · 2026-10-06
- CutBCE: TPU Kernel Eliminates OOM in Large-Vocabulary Recommendation Training, 91.9% Faster — _reachsumit · 2026-10-06
- Domain specialists orchestrated by a general model: a local-LLM architecture pitch for 8-16GB GPUs — CyberExplore · 2026-10-06
- Run a 37GB Qwen MoE in a browser tab: LocalMind streams expert weights from disk, matches llama.cpp output — naklitechie · 2026-10-06