Implementing 3D Finite Difference Codes on the GPU
Implementing 3D Finite Difference Codes on the GPU
Implementing 3D Finite Difference Codes on the GPU
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
C<strong>on</strong>clusti<strong>on</strong>s<br />
• <strong>GPU</strong> achieves high throughput for <str<strong>on</strong>g>3D</str<strong>on</strong>g>FD over regular grids<br />
– Bandwidth and instructi<strong>on</strong> rate advantage over CPUs<br />
• Multi-<strong>GPU</strong> <str<strong>on</strong>g>3D</str<strong>on</strong>g>FD performance scales linearly<br />
– For large enough problems, communicati<strong>on</strong> is hidden by computati<strong>on</strong><br />
<strong>on</strong> internal data<br />
– “large enough” depends <strong>on</strong> stencil size and cluster c<strong>on</strong>figurati<strong>on</strong><br />
• Number of network and PCIe transfers<br />
• Multi-<strong>GPU</strong> scaling can be reliably modeled<br />
– Using throughput rates for <strong>the</strong> <strong>GPU</strong>, network, and PCIe bus