The following chart shows Microwulf's performance
on HPL's WR00R2R4 operation with PxQ = 2x4, NB = 160,
and increasing values of N:
We tried values of N above 30,000, but these produced "out of memory" failures. The curve clearly shows the performance starting to plateau, but we think we could have eked out more flops, if we had more than 1 GB RAM per core. However, as it is, Microwulf's 8 GB of RAM consumed over 40% of our budget, so more RAM would most likely increase the price more than it would increase the performance...
Joel Adams > Research > Microwulf > Performance > HPL