Re: Why Alpha? (aka: is it worth it?)

["Adam C. Powell, IV" <adam powell nist gov>]

"Daniel J. Frasnelli" wrote:

> > One tiny correction: including load/store, alpha 21164 can execute up
> > to *four* floating-point instructions per cycle.  Which is how Goto's
>         Hmm. Have a reference for this? The 500Mhz 21164a is rated for
> close to one GFLOPS per second per processor, which I thought could be
> explained by one FADD, one FMUL per clock cycle (0.5Gc*2).

Right, 2 FLOPS = floating point *operations* per clock, i.e. one add and one
multiply.  But if this were all you could do, you'd real fast run out of
registers to add and multiply together, and would not get anywhere near the
theoretical 2 operations/clock on a sustained basis.  So, four *instructions*
per clock refers to these two plus two loads and/or stores to feed the
registers, which don't count in FLOPS numbers.  The assembly source of K.
Goto's gemm is a great place to see this in action- the inner loop comes very
close to 2 FLOPS/clock, and the whole thing sustains an average of about
1.5-1.6 FLOPS/clock, which is very good.  (I think it's slightly faster than
the DU gemm, but don't quote me on that.)

Alpha is pretty amazing, isn't it?

> > Also, Daniel, does your high regard for SGI Origin take price into
> > consideration?  We've found that Origin nodes cost roughly five times
> > as much as
>         Eh, sir? Please re-read my message.  My "platform of choice" is
> constrained by what we have available - an 8 proc Origin2k and 4 proc
> Onyx2 RE2.

Sorry, I misunderstood. :-)

Zeen,

-Adam `Cold Fusion' Powell, IV http://www.ctcms.nist.gov/~powell/ ____
USDoC, National Institute of Standards & Technology (NIST)  |\ ||<  |
Center for Theoretical and Computational Materials Science  | \||_> |