> As Samuel Lipoff and Aaron Blosser have reported, they see some measurable
> performance improvement when comparing XEON CPUs with plain vanilla PIIs. I
What they appear to be reporting is a decrease in the iteration time
reported by Prime95. This comes from dividing the change in the
CPU clock count register by the program's estimate of the
processor speed. If the speed is too high, then the reported
iteration time will be too low, by the same proportion.
I think the Xeon performance is being overrated.
I suggest a simple test which will only take a few minutes to
execute. Put Prime95 into Advanced mode (if you don't already
have it that way) and time execution of testing 2^11213-1 using a
stop watch. I reckon the times for a Xeon @ X MHz and a PII @ X
MHz will be very close - whatever the reported iteration time says.
George - wish list - estimate Xeon processor speed correctly!
> have a PII Overdrive, which is basically a PII XEON in a different package
> such that it fits socket8 (PPro) motherboards, i. e. this CPU has 512KB L2
> cache running at core speed and a deschutes core, just like the PII XEON.
> However, it runs only at 66*5 = 333MHz.
> On my machine (dual board, 1 CPU, 440FX chipset, 256MB ECC EDO) my
> "RollingAverage" value is constantly around 850 (before the upgrade, this
> number was usually around 970), even if I leave the machine mostly alone (I
> have not done any precise measurement of "time/iteration" though). That is
> despite the fact that - as others have seen in the case of the XEON CPUs -
> the setup program overestimates the clock rate of the CPU (in my case only
> by a mere 10%, however, if memory serves). Any ideas where the difference
> comes from? Poor memory performance of the FX chipset? Penalty for extra
> writes due to ECC?
Same chipset so this isn't the problem. There is no penalty for
accessing ECC unless the (hopefully extremely rare) event of
actually having to correct a bit error occurs, in this case this
memory cycle will be extended by a few clocks.
No. If your CPU speed is estimated high by 10%, then your
performance index should be low by 10%, since the "rolling
average" performance index is calculated using the time difference
between iterations 64K apart, using the time-of-day clock, not the
CPU cycle counter.
The rest of the difference is accounted for by the fact that a CPU
with a higher multiplier always takes more clock cycles to do the
same task, if there are any cache misses at all - you haven't
upgraded the memory bus bandwidth. e.g. my Celeron 266 system
takes approx. 90.5 million cycles per iteration on an exponent
around 6.6 million (384K FFT size), my PII-333 takes about 98
million cycles per iteration for a similar size exponent; both are
running on a 66 MHz bus. Or, in "rolling average" terms, the C266
manages about 1280, the PII-333 about 1240. (Nevertheless the PII-
333 iteration time is about 0.297, whereas the C266 is around
0.340, so the PII-333 is definitely faster!)
As the multiplier goes up, so the efficiency tends to fall.
> BTW, has anybody any performance values of Celeron300A CPUs running at
> 450MHz? In general, given the same L2-cache speed, how much of an impact
> would the smaller size (128KB vs. 512KB in the XEONs) have?
I get the impression - though I have no direct evidence - that the L2
cache size does not have a major performance effect so far as
Prime95 is concerned - even 512KB cache is nowhere near big
enough!!! (The Xeon with 2MB cache may be a few percent better,
but you could buy a lot more power in Celeron systems for the
price of just one 2MB cache Xeon processor)
In fact I have the distinct impression that (for Prime95, and for
components up to 333 MHz) a Celeron with 128K cache is actually
marginally quicker than a PII at the same bus speed & multiplier.
I'd expect the Celeron 366 and 400 to perform not quite as well as
you'd expect from the C333 performance, certainly a PII-400
(4x100) should outperform a C400 (6x66) fairly comfortably, even
for Prime95. Nevertheless, in price/performance terms, the
Celerons look extraordinarily good. The real downside is no SMP
support...
Also I don't condone overclocking ... personally I'd rather produce
two good results than three shaky ones ... however, if you *must*
do it, I'd reccomend using one of the motherboards with
intermediate speeds (75 & 83) instead of just 66 & 100, so you can
step back from +50% to a more moderate value & (hopefully) still
get reliable operation.
NB I'm *not* trashing Xeon - I'm quite sure that, for general mid-
range server applications, the Xeon will outperform PII comfortably -
especially as it supports up to 8X SMP, whereas the PII SMP is
limited to 2X. (But I'm glad I don't have to pay for 8 x Xeons!)
Regards
Brian Beesley
