[rescue] ooh, pretty

[James Lockwood]

On Tue, 19 Nov 2002, Mike Meredith wrote:

> Don't fall into the trap of assuming Mhz=speed (Dave obviously knows
> this) ... a quote from a supercomputing centre looking at the R8000
> said that it was half the clock speed of the R4400 but 4 times the
> speed (probably just looking at fp perf).

This is reasonably believable for the nicer R8k processors found in
various Challenge configurations.  The 75MHz R8k used in the I2 is
crippled with a small cache and less bus bandwidth.  The real reason for
its production was marketing, SGI was waiting for the R10k and couldn't
afford to have FP intensive customers jump ship.  It was useful for
keeping the FP crown but poorly balanced.

A 75MHz R8k I2 should not be faster than any dual R10k Octane unless there
is considerable disparity between other portions of the systems.  The
Octane has greatly improved bus technology and even a relatively slow R10k
in an Octane will outperform the R8k.

CPU to memory is _the_ major bottleneck in the I2, it's nearly equalled by
a SS5/110 and trounced by a U1/140.  The Challenges are 1.5-2x better (for
a single CPU, aggregate bandwidth with many CPUs is much higher), the
Octanes are 2-4x better.

-James