AMD hipCIM accelerates MONAI digital-pathology pipelines 10× on Instinct MI355X GPUs

AMD has published benchmark results showing hipCIM 26.06.00, its GPU-accelerated whole-slide image I/O library, integrated into MONAI's digital-pathology pipeline through the familiar from cucim import CuImage API path. Running on AMD Instinct MI355X with ROCm 10.0, hipCIM delivers approximately 10× faster GPU patch reading compared to OpenSlide, which translates to a 4.2× speedup in MONAI's PatchWSIDataset data loader and 1.6–3.8× faster end-to-end inference across three whole-slide images. The implementation is available through AMD's fork of MONAI (AMD-Ecosystem/MONAI), with changes fed back upstream, and the packages are distributed as amd-hipcim and amd-monai on the AMD ROCm package index.

The work targets a critical bottleneck in medical AI: whole-slide images are gigapixel pyramids of JPEG-compressed tiles, and the reader that decodes disk-resident slides into GPU tensors dominates wall-clock time. AMD's approach preserves numerical fidelity—rocJPEG decode matches CPU libjpeg-turbo to a maximum absolute difference of 1 on 8-bit values—while adding new GPU-resident 3-D volume reading capabilities for NIfTI and DICOM formats via cumed. The benchmarks used the pathology_tumor_detection bundle from the MONAI Model Zoo, a representative metastasis-detection workload on Camelyon16 H&E stained slides.

hipCIM integration architecture and benchmark methodology

• API compatibility layer: hipCIM exposes the identical from cucim import CuImage API that MONAI's WSIReader already uses for NVIDIA's cuCIM backend, enabling drop-in replacement without application code changes on AMD ROCm systems.
• ROCm 10.0 and MI355X stack: The benchmarks run on AMD Instinct MI355X (gfx950) with ROCm 10.0, Python 3.12, and a ROCm build of PyTorch; the loader comparison was additionally validated on MI350X with the same ROCm version.
• Patch-read throughput: hipCIM achieves 9,661–9,930 patches per second on CMU-1/2/3.svs slides versus ~943–951 patches/s for OpenSlide, with the speedup originating from GPU-resident JPEG decode via rocJPEG rather than CPU libjpeg-turbo.
• End-to-end inference gains: The pathology_tumor_detection bundle (TorchVision ResNet-18 fully-convolutional classifier, 224×224 patches at pyramid level 0, read as 768×768 regions split 3×3) shows 1.6–3.8× faster inference depending on the slide, with batch time stabilizing at ~0.13 s/batch versus 0.21–0.51 s/batch for the CPU reader path.
• 3-D medical imaging extension: hipCIM adds GPU-resident NIfTI and DICOM volume reading through cumed, a capability not previously available in the MONAI OpenSlide path, expanding beyond 2-D whole-slide pathology into volumetric CT and MRI workflows.

Why it matters

This release strengthens AMD's position in open-source medical AI infrastructure by demonstrating production-viable GPU acceleration on ROCm for a domain historically dominated by CUDA-based cuCIM. For Linux-based research computing clusters and hospital systems running MONAI on AMD Instinct hardware, the 10× reader speedup directly addresses the I/O bottleneck that typically wastes GPU cycles in digital-pathology inference pipelines. The upstream contribution model—AMD maintains a fork with ROCm-specific changes fed back to Project-MONAI/MONAI as pull requests—offers a path toward mainline integration that would reduce vendor-specific packaging fragmentation for distribution maintainers and clinical deployment engineers.