White Papers

Extreme GPU Computing
Page 8
4. Performance characterization
4.1 Bandwidths between CPU to GPU
We measure the host-to-device (H2D) and device-to-host (D2H) bandwidth of the five C4130
configurations. Figure 6 and 7 show the measured bandwidths. Two CPUs and eight GPUs (two internal
GPUs per K80 board) yield 16 CPU-to-GPU combinations. The CPU (host) and GPU (device) bandwidth
measurements for each configuration is shown in figures below. The Host to Device (H2D) and Device to
Host measurements are about 12000 MB/s, which are the state-of-art achievable in Gen PCIe, with a
peak of 15754 MB/s. It is noteworthy that the measurement is consistent and does not vary significantly
with configurations.
Figure 6: Measured host-to-device (H2D) bandwidths on various configurations
Figure 7: Measured device-to-host (D2H) bandwidth on various configurations
Four K80 Boards
Two K80 Boards
Four K80 Boards
Two K80 Boards