Performance of different GPU data access implementation strategies

The performance tests were conducted using two NVIDIA V100 GPU cards interconnected via NVLink interconnect. The hardware environment was provided by the imec GPULab.

The resulting performance data are provided to serve as a valuable reference for researchers and practitioners interested in designing multi-GPU applications or algorithms. Using a simple GPU kernel function computing the element-wise product of two vectors, we compared the GPU execution times obtained with traditional explicit host-GPU data transfer, and with Unified Memory and Unified Virtual Address mechanism-based implementations.