CPU benchmark advice ??

[Bohdy]

Hi, I'm looking for a CPU benchmark that is a good indicator of the the actual processing power of the CPU, not just some meaningless number.

Optimally it would be easy to compare between architectures, and run without problems on a PPC processor.

It would be great to have the code in C or PPC assembly.

Any ideas?

 

[Saem]

What sort of apps, if it's SpecCPU is good so long as the problem domain(s) covered by it overlap with the problem domain that you wish to address.

 

[Bohdy]

Hmm, thanks for the suggestion, but that would not be entirely suitable for my application.

I need something even more portable. Much smaller requirements would be good.

I have heard of Dhrystone and Whetstone, how are these regarded?

Any other suggestons would be welcome

 

[Saem]

SpecCPU is about as protable as it gets, AFAIK.

You might want to look around for some straight up ANSI C stuff. You could compile them with GCC so they'd have roughly equivalent compilers.

There are also games if you're interested. Specifically Quake 3 or RTCW, if you want something free and fairly current, RtCW:ET is a free download. THe problem is that it spends ~70% of it's time in the OGL driver -- extrapolating from Quake 3.

AFAIK, whetstone and drystone are pretty useless.

There is also a question of what's your problem domain? What specific sub-systems of the CPU do you want information about? Is it mainly integer, FPU... and then is it latency, throughput or both?

 

[Bohdy]

I am looking to bench the whole thing pretty much, int, fpu, memory and cache performance.

There is also a SIMD extention, but I doubt I will find something to bench that without customisation.

I also need something will a rather small memory footprint. Around 2 meg if possible.

To be specific, it is a custom ppc750cx chip... ;-)

 

[Saem]

Neat find.

Well there is the entire Spec suite which measures all sorts of stuff. The problem is IIRC, it costs money. =(

I'm not sure if there is some sort of free version of it. One thing you could try is to find benchmarks from the Linux community. I believe you can use compile times as one bench, you'll need to figure out what parts of the system are stressed. There also might be some cross platform applications such MPEG X encoders which could give you more information about bandwidth and some processing capabilities.

I don't think you can find individual sub-system tests nor should you, really. I think from an application point of you, you're going to be testing a collection of sub-systems.

 

[AAlcHemY]

Why not Sisoft Sandra ? :?

 

[Tim Murray]

Why not Sisoft Sandra ? :?

Read the thread. It's a custom chip, and he needs it to be small.

It ain't a Pentium 4 he's talking about.

Prime95 came to mind, but it seems to only be distributed in binaries. There's a version called Glucas which seems like it might be portable to your plattform-- http://sourceforge.net/projects/glucas .

 

[Dave H]

Unfortunately, most benchmarks that are simple and available as source are synthetics. If you're interested in memory performance, cache hierarchy, and best-case FPU throughput, Linpack would be a decent choice. Along the same lines, benchmarks of the various BLAS libraries, particularly if you'll be doing anything involving linear algebra.

SPEC is out of the question if you want to fit into such a small footprint. Generally speaking, benchmarks for embedded systems are pretty rudimentary; oftentimes Dhrystone/Whetstone are quoted, which is worse than worthless. There are a few decent suites emerging, though. The new H&P uses EEMBC to discuss embedded performance; there's even a result up at the site for the 750cx. From a quick perusal it appears you might need to become a member (and presumably pay money) to get at the benchmark yourself, but the score report is pretty detailed so that might tell you a bit of what you want to know.

EDIT--just to be clear, that's the EEMBC site I'm referring to above

 

[Bohdy]

Thanks a lot for all the advice

It could be a little while before I try them all however.

Cheers again !