[rescue] RE: SGI vs. Sun

[Janet L. Campbell]

On Thu, 13 Jan 2005, Carl R. Friend wrote:

>    On Thu, 13 Jan 2005, Janet L. Campbell wrote:
> 
> > We had a dual SM81 SS20 here for a while, in aggregate CPU power it
> > was comparable to an Indigo2 R10k/200MHz, and more capable I/O-wise
> > (4 FWD  SCSI channels, hme, and 4MB of PrestoServe at its peak).
> 
>    Where in this hierarchy does the R8k Indigo^2 fit?  I've got one
> in my collection and am curious.

The R8k I2 was an odd duck...it only ran at 75MHz, and had less cache than 
the R8k's used in the Power Challenge/Onyx.  It was basically intended to 
give SGI back the FP crown while they waited for the R10k to be finished.

It's difficult (some would say "frustrating") to optimize for.  Code that
can take advantage of its multiissue pipeline can perform very well,
attention to coding techniques and good compiler optimizations are
important.  Handcoded libraries (like SGI's BLAS implementation) were
needed to wring out maximal performance.  I seem to remember it being
roughly comparable to a mid to high range I2 R4400 for general non-FP
work.

The R10k is a cleaner design, scaled better, and performed much better 
under suboptimal (read: "most") conditions.  Optimizations that help the 
R8k also help the R10k, but the difference between best and worst case is 
much smaller.  A useful parallel might be R8k::R10k = i860XR::i860XP.

-Janet