Computational and Memory Performance

Memory Performance

• Again, the computational performance scales with memory performance

• MSP shows significantly better memory throughput over autotasked code

Floating-Point Performance

• Larger apparent cache yields greater than 4x speedup for MSP executable

• Average vector lengths: T90: 99.58 SV1: 60.53

Previous slide

Next slide

Back to first slide

View graphic version

• We can now add two values to our floating point results

• 1) Autotasked and run with 4 cpus (not optimized)

• 2) Multi-streamed and run on an MSP

• Note the more than a factor of four speedup for the MSP case.

• - larger effective cache & more memory bandwidth

• Performance of auto-tasked case could be improved using compiler directives

• -Notice that as with the SSP runs, floating point performance and memory bandwidth scale together