Interesting quote from JC...
"he K7 has an immense floating point unit, I'm running it through paper simulations
against the P3 currently, and it looks like we'll get a serious benefit of the K7 over the P3 (it's much
more than a clock speed faster in fpu from the looks of it), especially since the internal scheduler is so
damn big, a great boon for o-o-o execution. The integer unit is also amazing (three issue with better
caching, compared to the generally dual issue of the P2.
Like I said, I'm running simulations. With 1:1 alternating multiply/add instructions, the K7's fpu can do
twice the rate of the P3. Likewise for all multiplies, though all adds would be about equal speed for P3
and K7. Those matrix operations for 3D processing is more of a mix of the two. In trancendental (FSIN,
FCOS, etc) and complex fpu code (FSQRT, FDIV), the K7 has much lower latencies and pretty low
throughput (at least on the sqrt and div) compared to the P3, so high end engineering stuff may just
kick ass on the K7 over, say, the Xeon. I'll post up numbers on my site if I discover anything.
BTW, I don't mean to imply the K7 will run twice as fast as the P3. That's peak performance levels, which
is affected by logic & integer instructions (which are always intermingled) as well as caching issues, which
I'm not qualified to handle."....
Jim |
