Time for the true: L3 cache test.

[MacTron]

L3 caché it's very important, especially in the PPC era where the CPU speed was increased a lot more than system bus speed. Even true that everybody claims that 2 MB of L3 is better than only 1 MB ... but how much?

I have changed my 1.5 Ghz MDD cpu with 1MB of L3 cache by a 1.5 Ghz cpu with 2MB of L3 cache from a xServe ( both CPUs overclocked at 1.5 Ghz) and make some test before and after the change. So being exactly the same machine, if some differences appear, the cause will be the amount of L3 caché.

MacBench 3 only shows a 10% of speed increase on hardisk related test. Cinebench 2003 show a suppressive 50%  of speed increase on the rendering test ( something went wrong in this test I think ) , and a 10 % of speed increase in the Shading test.

The most important test was the video encoding test, because it uses a lot of CPU and memory bandwidth. I had taken the average of encoding 4 times the same video before and after the CPU change and comparing the average encoding speed in both cases the result was a 5% faster using the 2MB L3 CPU.

[DieHard]

Thanks for the results... in a DAW environment, I can "feel" the difference quite dramatically with G4s that have less than 1 MB Cache.

So, I would say, another good test is the CPU meter in the DAW with the same project loaded.

[MacTron]

So, I would say, another good test is the CPU meter in the DAW with the same project loaded.

Yes it is. But take into account that this kind of test is very hard to perform. I have tried it in other occasions, (this is my second MDD/xServe    ) a 5-10% it's hard to detect with the Cubase CPU meter, and if you use a very complex setup to maxed out the CPU, over a 70-80% the meter can easily goes to 100%. You'll become crazy until you get a trusted result.
But yes, You can "feel" the difference ...

[3rd Degree]

Don't forget that L3 cache is a larger, slower version of L2 cache. Maybe compare the L2 cache sizes of the processors to get a better picture of relative performance.

Back in the day the test we used to do was the 'platinum reverb' test. Basically load up an empty Logic file and add platinum reverbs until the system overloaded.

A great intro on tweaking MacOS 9 era Macs for audio performance (compared to DTP performance) can be found here:

A Philosophy of Tweaking
http://www.xlr8yourmac.com/audio/issue67.html

[MacTron]

Don't forget that L3 cache is a larger, slower version of L2 cache.

... and L2 cache is a larger, slower version of L1 cache

The object of the test is to measure the speed differences between two identical MDD CPU modules with identical CPU (PPC 7455), identical L1 and L2 caches, but different amount L3 cache

Maybe compare the L2 cache sizes of the processors to get a better picture of relative performance.

Two CPUs with different amount of L2 cache, are two different CPUs like 7447 (512 kB of L2) and 7448 (1Mb of L2), at least in the 68k and PPCs used in Macintosh. And yes, they have different performance, due to the cache size and CPU capabilities.

[3rd Degree]

Yep - got it in one Mactron. I couldn't tell from your first post whether or not you were testing processors with the same L1 and L2 caches, which is why I brought it up.

I'm also dubious as to the value of CPU meters as an accurate gauge of performance across systems. But settling on a common testing platform and plugin, like Logic and Platinum Reverb, is a good way of simulating real world conditions. Given that logic plugins are in the app itself, it would be quite easy to create an archive containing the app, dongle emulator, a file loaded with a few platinum reverbs and instructions for non-logic users on how to use the archive to test the performance of their machine.

Reporting the number of plugins open before the system started overloading would highlight the benefits or otherwise of higher clock speeds, dual processors, more cache etc. This obviously would be testing processing power rather than disk drive performance.

I'm happy to pull this together if folks are interested.