Benchmark

[Kiko]

Hi there,


We just finished a benchmark of some RISC processors which includes the
IBM RISC S/6000 Model 530 & 320, a DEC Station 5000 Model 200, 
a SUN 4 Model 330, a Clone PC 386/33Mhz (w 80387) and SUN 3 with FPA.

The goal was to determine the speed-up achieved by the IBMs and DEC over the
SUNs when running totally CPU bound processes. The 386 and SUN 3 were included
as a curiosity only.

We used sample problems in the domain of Neural Networks, namely some
Back Propagation Network Training & Simulated Annealing Combinatorial
Optimization. Both problems require a great deal of floating point processing.
Here goes the numbers:

The S/6000 Model 530 scored an average of 4.5 faster than a SUN 4 Model 330.
The speed-ups ranged from a low of 3.7 to a high of 5.29 the processing speed
of the SUN 4. It was the clear winner of the benchmark.

The second place was for the S/6000 Model 320 which averaged 3.14 times faster
than the SUN 4. The speed ups ranged from a low of 1.88 to a high of 4.10.

In third came the DEC Station 5000 Model 200. It was a little disappointing
given that this is the high end model of the new workstation line. We were
expecting numbers closer to the S/6000 Model 530. In any event, the DEC
scored an average speed-up of 2.16 times the speed of the SUN 4. The speed-ups
ranged from a low of 1.73 to a high of 3.20, but always below those of the
S/6000 Model 320 when running the same programs. It is important to note that
the processing power of the DEC gets very close to that of the IBM 320 when
the size of the problem running on both machines increase beyond a certain
point. This probably has to do with the smaller data cache on the IBM. 
Nevertheless, the IBM always scored higher.

The 386/33Mhz scored a pretty constant 0.25 times the speed of the SUN 4, 
well above the 0.025 scored by the SUN 3. It was really interesting to note
the 10-fold speed-up of the 386 against the SUN-3 achieved by such inexpensive
machine. And even more interesting the 100-fold speed-up achieved in just
two generations, from the SUN 3 to the newer RISC processors.

We are reading a report which can be obtained upon request.
Send mail to:

     camargo at cs.columbia.edu

     Francisco A. Camargo
     DCC Lab, Columbia University
     511 Computer Science Department
     New York, NY, 10027


/K.