Summary
This paper presents Futurism, a heterogeneous K-means clustering framework for mobile platforms operating under tight power and performance constraints. While GPU-accelerated implementations offer substantial parallelism, they can leave CPUs underutilized and spend significant time stalled on memory accesses. Earlier CPU-GPU approaches use both processors but rely mainly on scalar CPU execution, leaving the CPU’s SIMD capabilities underused.
Key techniques include:
- Using SIMD instructions for distance computation on the CPU
- Organizing data in a sample-wise layout to improve CPU-side execution
- Dynamically distributing workloads across CPUs and GPUs to improve resource utilization
Together, these optimizations deliver a 10% end-to-end speedup over prior CPU-GPU heterogeneous implementations on NVIDIA Jetson Orin AGX.
Download & Original
- IEEE Xplore page: https://ieeexplore.ieee.org/document/11690603
Copyright note
The full text of the paper is available from IEEE Xplore and is subject to IEEE copyright. This page provides an original English summary and metadata only; consult the IEEE PDF for the complete article.