Mercury Cell presentation (new performance comparison with FFTs, image filters)

[Titanio]

I think the reason for no Intel numbers in those image filter tests is that Mercury doesn't really work with Intels. The FFT numbers for the Intels were gathered from that FFT site. Mercury only benched on the PPC, Cell and Freescale themselves. Maybe if there was a site out there with image filter performance numbers for various chips, including Intels - like for the FFTs - they'd have included those numbers.

 

[Shifty Geezer]

And that's in a highly parallelisable non branchy unrollable predictable workload that should be the Cell's forté.

Perhaps, but that's not on floating-point intensive operations. These matrices are pretty much n*n FMADDs on n*n data components, or one FMADD per pixel. It's not really the multiple hardcore vectorised float operations the SPEs are ideally designed for, and not where its advantages will shine.