DSPC-8681 Benchmarking

Dorian Fernandez1

Other Parts Discussed in Thread: FFTLIB

Hello,

I'm currently working on running benchmarks on the DSPC-8681 card and I came across the following document that has benchmarking numbers for FFT and SGEMM. I was wondering what code was used for these benchmarks and what optimizations were done in order to achieve those results (EDMA, OpenMP, etc.).

Link to document : http://www.farnell.com/datasheets/1737036.pdf

Thanks,

Dorian

over 11 years ago

0 lding over 11 years ago

TI__Guru* 95265 points

Dorian,

We are checking with marketing team about those numbers and will let you know once we have more info.

Regards, Eric

0 lding over 11 years ago in reply to lding

TI__Guru* 95265 points

Dorian,

The number may came from DSPLIB or FFTLIB. The best FFT: 4096 points takes 27502 cycles from L2. The Gflop calculation is as following: 5*4096*log2(4096)/27502*1.2G=10.7Gflops. If the paper is quoting 8-core performance, it could be DSPLIB FFT, which is a little slower.

SGEMM performs at 10Gflops per core in BLAS. The paper could be DSPLIB again, which also offers SGEMM kernel and again a little slower.

Regards, Eric

0 Dorian Fernandez1 over 11 years ago in reply to lding

Prodigy 20 points

Eric,

Thanks for the help!

-Dorian

Processors

Processors forum

DSPC-8681 Benchmarking