The Case for Higher Computational Density in the Memory-Bound FDTD Method within Multicore Environments

<table class="table-group" id="tab4"><tr><td><table class="table"><tr><td class="thead-hr" colspan="3"><hr/></td></tr><tr class="thead"><td align="left">Algorithm</td><td align="center">Throughput (MCells/s)</td><td align="center">FDTD/Algorithm throughput</td></tr><tr><td class="thead-hr" colspan="3"><hr/></td></tr><tr><td align="left">FDTD</td><td align="center">383</td><td align="center">1</td></tr><tr><td align="left">FV24</td><td align="center">159</td><td align="center">2.4</td></tr><tr><td align="left">FV24-WB</td><td align="center">198</td><td align="center">1.9</td></tr><tr><td align="left">S<svg height="15.05" id="M31" style="vertical-align:-2.30685pt" version="1.1" viewbox="0 0 34.950001 15.05" width="34.950001" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink">
<g transform="matrix(.017,0,0,-.017,.062,12.113)"><path d="M293 -169q-145 0 -145 128q0 35 4 106t3 104q0 17 -2 28t-10 25.5t-28 23t-51 10.5v29q31 2 51 10.5t28 23t10 25.5t2 28q0 31 -4 101t-3 104q0 132 145 132v-28q-51 -7 -67.5 -30t-16.5 -63q0 -32 6.5 -99.5t6.5 -100.5q0 -97 -83 -115v-4q83 -20 83 -117
q0 -34 -6.5 -99.5t-6.5 -96.5q0 -41 16.5 -65.5t67.5 -31.5v-28z" id="x7B"></path></g><g transform="matrix(.017,0,0,-.017,5.961,12.113)"><path d="M412 140l28 -9q0 -2 -35 -131h-373v23q112 112 161 170q59 70 92 127t33 115q0 63 -31 98t-86 35q-75 0 -137 -93l-22 20l57 81q55 59 135 59q69 0 118.5 -46.5t49.5 -122.5q0 -62 -29.5 -114t-102.5 -130l-141 -149h186q42 0 58.5 10.5t38.5 56.5z" id="x32"></path></g><g transform="matrix(.017,0,0,-.017,14.12,12.113)"><path d="M95 130q31 0 61 -30t30 -78q0 -53 -38 -87.5t-93 -51.5l-11 29q77 31 77 85q0 26 -17.5 43t-44.5 24q-4 0 -8.5 6.5t-4.5 17.5q0 18 15 30t34 12z" id="x2C"></path></g><g transform="matrix(.017,0,0,-.017,20.818,12.113)"><path d="M412 140l28 -9q0 -2 -35 -131h-373v23q112 112 161 170q59 70 92 127t33 115q0 63 -31 98t-86 35q-75 0 -137 -93l-22 20l57 81q55 59 135 59q69 0 118.5 -46.5t49.5 -122.5q0 -62 -29.5 -114t-102.5 -130l-141 -149h186q42 0 58.5 10.5t38.5 56.5z" id="x32"></path></g><g transform="matrix(.017,0,0,-.017,28.977,12.113)"><path d="M283 255q-31 -2 -51 -10.5t-28 -23.5t-10 -27t-2 -30q0 -28 4 -97.5t4 -107.5q0 -128 -146 -128v28q50 7 67.5 31t17.5 63q0 33 -7 100t-7 98q0 96 84 117v4q-84 20 -84 115q0 34 7 102.5t7 98.5q0 40 -17.5 63t-67.5 30v28q28 0 50 -4t45.5 -16.5t37 -40t13.5 -68.5
q0 -37 -4 -106.5t-4 -98.5q0 -18 2 -30t10 -27t28 -23.5t51 -10.5v-29z" id="x7D"></path></g>
</svg></td><td align="center">144</td><td align="center">2.7</td></tr><tr><td align="left">S44</td><td align="center">215</td><td align="center">1.8</td></tr><tr class="table-tr"><td colspan="3"><hr class="tbody-hr"/></td></tr></table></td></tr></table>

Comparison of GPU kernels’ throughputs for the various FDTD algorithms.

International Journal of Antennas and Propagation

tab4

Table 4

Table 4: The Case for Higher Computational Density in the Memory-Bound FDTD Method within Multicore Environments 