1 Apr
2015
1 Apr
'15
5:04 p.m.
From: Sean Caron <scaron at umich.edu>
You are right, even the interactive performance of the
R8K had me a bit
taken aback...But I understand in its day it really flew on the FP codes;
Yup. The R8000 was pretty specifically designed to be a desktop super, for
some value of "super". And for 1995, it was pretty remarkable. It was the
first superscaler MIPS, and the dedicated FP unit was pretty much pushing
the envelope on single chip tech. But integer performance wasn't a
priority, and you could tell.
I don't know if this claim was ever really
substantiated in the real world
The MIPSPro compilers usually did a really good job of optimizing for the
R8000, and for the jobs it was designed for (non-vectorizable floating
point) it really flew (300ish MFLOPS...which for 1995 was nothing short of
amazing in a desktop).
I would note, however, that the common comparison to the Y/MP is pretty
much pure marketing. The Y/MP and the R8K were on-par for non-vector FP
problems, but as soon as the compiler could substantially vectorize the
problem, the Y/MP could be many times faster than the R8K for the same
codebase.
In the end, we found the R8K was a nice dev machine (due to a good bit of
source compatibility between the MIPSPro and Cray compilers), but not
generally cost effective, and pretty much outmatched all around by the R10k
and later.
But it's a neat little oddball
Very.
KJ