🚀 go-pugleaf

In article <dhess.728160535@Xenon.Stanford.EDU> dhess@Xenon.Stanford.EDU (Drew William Hess) writes:
>The San Jose Mercury News had a small article today re: SGI's announcement
>of their new supercomputers.  They also mentioned that the Power Challenge
>systems use the TFP, which is, I believe, based on an earlier MIPS design
>(R4000, I suppose) but with a floating point unit on steroids.  Has SGI
>made any formal announcement of the TFP?  I believe that SGI did most of
>the development on the TFP in-house before acquiring MIPS, so is SGI/MIPS
>going to make it available to third parties as they do with the Rx000 
>series?

TFP is an entirely new design.
It is not based on the R4000 design database.
A subsequent post gave the way to get more information on TFP, but I
thought I would clear up this mis-speculation here before it propagates.
For the curious, TFP is an acronym for "True Floating Point" --
or at least that is what they said at the announcement today.
The company talked about TFP last year when talking about our long-term
roadmap.  The TFP design was begun prior to the SGI/MIPS merger by SGI
to satisfy an SGI market need for higher performance FP than the R4000.

On the topic of FP units and steroid use...
The R4000/R4400 are noticeably different than the current
high-end HP, DEC, and IBM efforts in at least one way.
The R4x00 doesn't emphasize floating point to nearly the same degree.
The current "best chip" R4400 at 75 MHz in a relatively low cost
desktop system provides better "integer benchmark" performance than
anything HP, DEC, or IBM have announced with the exception of the
(much higher cost) 180MHz and 200MHz Alpha systems.
This isn't an accident.

Having noted that we fall (quite) short in some of the FP benchmarks...
Experience in the marketplace indicates that the single-value
summary of existing "FP benchmarks" like the two SPECfp numbers
fail to predict the actual relative performance of competing systems
for many programs that the users believe are "FP applications".
Sales/marketing folks at SGI say that Crimson boxes (50 MHz R4000) 
often do as well or better than fast IBM and HP boxes in customer
benchmark situations for "FP applications".  One factor may be that
a computer buyer runs problems/programs that are larger than benchmarks
and bust any of the primary caches.
Another factor seems to be that the customers are mostly interested in
the FP aspect of the applications but that these applications in fact
spend a lot of time in "integer" code.  Since the R4x00 is very good
at that it speeds that part of the program up and...
This isn't a surprise if you look at the detailed SPEC benchmark
results -- the FP benchmarks show much more variability among
processors than the integer benchmarks do.
Unsurprisingly, if you wanna know how fast a system really is for 
your exact application, you have to check it out.

If you are an FP multiply-add number head and you use integers for
addresses and do-loop indices only then you probably aren't really
excited about the R4x00.
Now you can be excited about TFP.

-- 
Charlie Price      cprice@mti.sgi.com        (415) 390-4457
MS 10U-178 / MIPS Technologies / P.O. Box 7311 / Mountain View, CA   94043