[rescue] R8k I2's
Francisco Javier Mesa-Martinez
lefa at ucsc.edu
Tue Jan 7 23:48:13 CST 2003
> OK, the R8k was a stop gap measure until the R10k was ready. SGI
> needed a machine with really good floating point performance. So, they
> introduced the R8k, which used some really cutting edge design ideas to
> get this performance. One of the problems is that they traded integer
> performance for floating point. So many applications that aren't
> floating point heavy ran slower on this than on previous chips. Also,
> the R8k had a deep pipeline and having to flush it was very expensive.
It was not the fact that it was deeply pipelined, but rather that it was
pretty wide issue scalar for the time. And it haf a lot of restrictions on
which instructions and how they could be paired in the parallel pipes, and
all the dependencies and stuff. As I mentioned in previous posts I had to
actually code for these things, thus my dislike for this particular
processor. The R10K made things much much much...... much much nicer. And
Fortran compilers are OK, since most of the code in fortran is loop
intesive anyways, so optimizations are pretty straight forward. Most of my
stuff was C and we had to do a lot of hand optimization, I still have the
R8K handbook around (it is really used!) and I am glad I have not opened
in a while, the only reason why my previous employer stuck with SGI was
all the effort that we put into getting the R8K code out, at the time was
the fastest FP machine out there. And the Alphas were not scaling as large
as the Power Challenges... plus I did not buy the machines, so we worked
with what we had.
The R8K worked best on the challenges since it came in the 90Mhz version
and 4MB of cache, which you really need for this processor. The I2 I used
at home (so I could code from home :-) ) had only 2MB and ran at 75Mhz.
The R10K is pretty much a better balanced single chip R8K, and it is a
great design in my opinion. And for most purposes is almost 2x as fast as
the R8K (in FP mind you). Also the r8K is a two chip processor, so
you could only fit half the number of r10Ks in the same chassis...
> Anyway, when the R8k was happy, it kept up with the later released R10ks
> fairly well.
Hum... as a person that used heavily both machines I find this hard to
believe. There are some hardcore r8K supported that like that whole myth
of the R8K running certain code faster than an R10K, and it was usually
examples of code that would fit in the 4MB cache for the R8K, wereas the
R10K only had 2MB of cache, so certain memory spills are to blame.
> Now, in my opinion, an R8k is in no way shape or form a waste of space.
> It might not make for the best interactive desktop workstation, but for
> running floating point heavy custom software, my understanding is that
> it is still very usefull. Especially when in the form of a Challenge,
> like the one that George has.
Well noone is claiming that the R8K is a waste of space, but if anyone
wants to spend $$$ on a machine for actual use, their bucks will be better
spent on an R10K machine. For the coolness factor, then yeah go for the
R8K but make sure that you understand that it will not be a responsive
workstation (when compared to more balanced machines).
Just the nightmares that machine gave me make me a bit passionate about
the R8K... aaahhh the ulcer! :-)
More information about the rescue