Eric wrote:
> My motherboard has two of those CPUs, and is a dual
> socket motherboard, would it be 402 points x 2 = 804 points.
> Thank you very much for suggestions
> Eric
>
I don't know enough about Passmark to say what the result would be.
Looking at the page here -
http://www.cpubenchmark.net/cpu_test_info.html
I would guess running the same test on each of the
CPUs, would result in some bus contention, especially
in the compression test. Some of the other benchmark
sites, see their highest dependence on memory bandwidth,
in compression type tests. So that probably won't double.
(I.e. Two processors can't do twice the work in that case,
because it takes longer to get data from memory.)
*******
If you want a means of estimating, you could use
Cinebench, run it on one processor, run it on two
processors, and use the ratio as a "scale factor"
to multiply by the 402 points number above.
http://http.maxon.net/pub/benchmarks/CINEBENCH_11.529.zip (145,723,303 bytes)
You unzip that, and run the "CINEBENCH Windows 32 Bit.exe"
executable. There is nothing to install. (Run the 64 bit
version if you have a 64 bit Windows OS.)
The first thing to do, is go to the File menu, and remove
the tick mark next to "Keep Best Score". You want the
test to update the result after each run, no matter whether
it is better or worse.
On my computer, without changing the preference item, the
program tells me I have a 2C/2T test situation. That is
two computing cores and two test threads. That will likely
be your default as well.
Running the test might take 5-15 minutes. Click the "Run" button
next to "CPU". Two windowed areas should open, showing the
progress of two test threads. When the work of a thread is
completed, another thread is dispatched. Two threads should be
"in flight" until the test is finished. The program activity
being computed, appears to be ray tracing in an image
with multiple light sources.
The program will report a score after the test is complete.
For example, with two test threads, my score was 1.78 points.
Next, go back to the File menu, and select Preferences. You
can increment or decrement the number of test threads. Change
the number to "1", tick the box for Custom Number Of Test
Threads. Once you've set that preference, click the run
button next to "CPU" again. When the program runs, you should
see just one windowed area appear at a time, implying one test
thread is being used. The test will take roughly double the
time to run to completion.
My result for the second run, using only one test thread (which
uses one computing core), my score was 0.90.
If I take the ratio of those two scores 1.78/0.90, the result
is 1.978 . That means my processor scales pretty well. It should
have executed two threads at twice the rate, but only managed
to do it 1.978 times as fast.
Take the ratio, using your dual socket motherboard, and see
how much of a bottleneck your shared bus is. Your ratio should be
some number less than 2. You can multiply that number by the
Passmark result of 402 if you like, as an estimate of your
dual socket result to be expected.
To give some other examples, if you use a Core2 Quad core in
a test like that, the internal shared bus is a bit of a bottle
neck, and running four test threads, runs at about 87% of the
speed you would have expected. So the bus structure costs you
about 13% of the theoretical performance. If you repeat the
test with a Phenom quad core, four cores runs at exactly
four times the speed of one core, telling you there is no
performance loss. I think the Core i7 is the same way as the
Phenom. It has an integrated memory controller and a relatively
large cache, and hides bus issues well. But the Core2 quad core
isn't quite as good. I'm expecting a similar effect on your
dual Xeon. It shouldn't be twice as fast with two test threads.
I think the shared bus scheme runs out of steam at around
four sockets. There would not be much point extending the
bus on your motherboard, and using five sockets, as the
bus would be choked.
I'm just guessing here, but your processors are likely
connected like this. The CPUs arbitrate to see who owns the
bus and gets to do a transaction with the Northbridge.
CPU0 CPU1
| |
+-----+-----+
|
Northbridge ----- Memory
An alternative means, is to do this. This approach is used
on more modern server motherboards, with colossal pin count
Northbridge chips. By using a private bus connection for each
processor, and perhaps hiding a snoop cache inside the Northbridge,
some of the bus bottleneck can be removed. It is my guess your
processor is a generation previous to this, and uses the first
figure.
CPU0 CPU1
| |
| |
| |
Northbridge ----- Memory
Have fun,
Paul
