What changed
AMD has integrated its hipCIM library into the MONAI framework, enabling faster processing of digital pathology workloads on AMD Instinct GPUs. The update, part of AMD Life Science 26.08, allows MONAI's WSIReader to leverage hipCIM through the familiar from cucim import CuImage API. This integration focuses on accelerating the data loading phase of the pipeline, which is critical for handling large whole-slide images.
What this means in practice
The integration of hipCIM into MONAI significantly improves the performance of digital pathology pipelines. On the AMD Instinct MI355X with ROCm 10.0, hipCIM achieves a ~10× speedup in GPU patch reading compared to OpenSlide. The MONAI data loader runs ~4.2× faster than the standard CPU reader on the CMU-1.svs slide. End-to-end inference across three whole-slide images is 1.6–3.8× faster. These improvements are attributed to the efficient handling of JPEG-compressed tiles and the reduction of host-copy overhead.
Limitations and availability
The current implementation of hipCIM in MONAI does not yet support GPU-resident processing for the entire pipeline. MONAI's GPU-resident path is not yet wired up, and the reader path still returns patches to the host. The results are based on the AMD Instinct MI355X with ROCm 10.0, and the benchmarks were conducted using Python 3.12 and a ROCm build of PyTorch. The hipCIM 26.06.00 and MONAI 1.6.0 versions are available on the AMD ROCm package index.
