high performance computing on graphics processing units: hgpu.org

The cost of data movement over the PCI Express bus is one of the biggest performance bottlenecks for accelerating data-intensive applications on traditional discrete GPU architectures. To address this bottleneck, AMD Fusion introduces a fused architecture that tightly integrates the CPU and GPU onto the same die and connects them with a high-speed, on-chip, memory controller. This novel architecture incorporates shared memory between the CPU and GPU, thus enabling several techniques for inter-device data transfer that are not available on discrete architectures. For instance, a kernel running on the GPU can now directly access a CPU-resident memory buffer and vice versa. In this paper, we seek to understand the implications of the fused architecture on CPU-GPU heterogeneous computing by systematically characterizing various memory-access techniques instantiated with diverse memory-bound kernels on the latest AMD Fusion system (i.e., Llano A8-3850). Our study reveals that the fused architecture is very promising for accelerating data-intensive applications on heterogeneous platforms in support of supercomputing.