HyPIM: Accelerating Hyperbolic Machine Learning via Processing-In-Memory
Published in 2nd IEEE Cross-Disciplinary Conference on Memory-Centric Computing (CCMCC 2026), 2026
Jinlin Wu, Ferit Ozdaban, Selvanathan Sadhasivam, Xiaoyang Lu, Yunhui Guo, Minxuan Zhou
Background
Hyperbolic machine learning uses non-Euclidean geometry to represent hierarchical data efficiently, benefiting applications such as graph learning and computer vision. However, existing GPU accelerators are primarily optimized for compute-intensive matrix operations and struggle with the memory-intensive transformations required by hyperbolic models.
In particular, Lorentz-model hyperbolic operations involve frequent data movement, element-wise computations, and reductions. These operations introduce a compute–memory imbalance and substantially underutilize GPU arithmetic resources.
Design
HyPIM introduces a GPU–processing-in-memory (PIM) heterogeneous architecture that accelerates hyperbolic machine learning by matching operations to the hardware best suited for them. Compute-intensive matrix operations remain on the GPU, while memory-bound hyperbolic operations are offloaded to HBM-based PIM units to reduce unnecessary data movement.
HyPIM combines near-bank processing units for vector operations and local reductions with channel-level units for transcendental operations. It further employs a hyperbolic-friendly data layout and in-flight wavefront scheduling to coordinate heterogeneous operations and improve hardware utilization.
Key Features
- Heterogeneous GPU-PIM Execution: Keeps compute-intensive operations on the GPU and offloads memory-bound hyperbolic operations to PIM.
- Specialized PIM Units: Uses bank- or bank-group-level units for vector operations and reductions, with channel-level units for transcendental functions.
- Data Layout Optimization: Organizes hyperbolic operands to exploit internal HBM bandwidth and reduce data movement.
- In-Flight Wavefront Scheduling: Coordinates different execution stages across PIM units to improve resource utilization.
Results
Experiments on hyperbolic machine learning workloads demonstrate:
- 3–7× speedup over an NVIDIA A100 GPU on evaluated HypFormer datasets.
- Approximately 30% average reduction in total energy consumption per inference task.
Conclusion
HyPIM demonstrates how hardware–software co-design and processing-in-memory can accelerate hyperbolic machine learning. By combining GPU computation with PIM-based execution of memory-intensive hyperbolic operations, optimized data layouts, and wavefront scheduling, HyPIM reduces data movement and improves the performance and energy efficiency of emerging AI workloads.