Genetic algorithms in software and in hardware-a performance analysis of workstation and custom computing machine implementations

The paper analyzes the performance differences found between the hardware and software versions of a genetic algorithm used to solve the travelling salesman problem. The hardware implementation requires 4 FPGA's on a Splash 2 board and runs at 11 MHz. The software implementation was written in C++ and executed on a 125 MHz HP PA-RISC workstation. The software run time was more than four times that of the hardware (up to 50 times as many cycles). The paper analyses the contribution made to this performance difference by the following hardware features: hard-wired control, custom address generation logic, memory hierarchy efficiency, and both fine- and course-grained parallelism. The results indicate that the major contributor to the hardware performance advantage is fine-grained parallelism-RTL-level parallelism due to operator pipelining. This alone accounts for as much as a 38X cycle-count reduction over the software in one section of the algorithm. The next major contributors include hard-wired control and custom address generation which account for as much as a 3X speedup in other sections of the algorithm. Finally, memory hierarchy inefficiencies in the software (cache misses and paging) and coarse-grained parallelism in the hardware are each shown to have lesser effect on the performance difference between the implementations.