Accelerating 3D convolution using streaming architectures on FPGAs

Different Precisions

As mentioned above, one of FPGA's advantages is the support for customizable number representations. Our previous work (Fu et al. 2008) has shown that, in certain cases of seismic computations, reduced precision provides equivalent results within acceptable tolerances. For FPGA designs, a reduced precision can significantly reduce the area cost and I/O bandwidth of the design, and multiply the performance with more computation units on the FPGA.

Figure 6 shows the performance we can achieve using a reduced floating-point precision. With a 16-bit floating-point precision, the multiple-stencil approach provides 49x speedup and the multiple-time-step approach provides 46x speedup.