This work-in-progress document presents preliminary Hwacha evaluation results. We first discuss the details of our evaluation framework. Using this framework, we compare a baseline Hwacha design to a Hwacha design with mixed-precision support to see how it affects performance, power/energy, and area. For the discussion, we name the Hwacha machine with mixed-precision support, highoccupancy vector lanes, or HOV. Obviously, this is not the only experiment we can run (consult the Hwacha microarchitecture manual for Hwacha parameters that we can change), however, it serves as a good example for the purpose of this document. We also validate the Hwacha design against the ARM Mali-T628 MP6 GPU by running a suite of microbenchmarks compiled from the same OpenCL source code using our custom LLVM-based scalarizing compiler and the ARM stock compiler.

EECS-2015-264