Synthetic Benchmarks be damned, my nMP just spanked my cMP.

[Macinsquatch]

So I have an 09 cMP that I have flashed and upgraded with dual X5690 CPU's, dual GPU's (AMD 7950 and Nvidia 970) that I intended to use as a workhorse for photogrammetry using photoscan pro.

Then I came across a deal I couldn't pass up on a nMP with the 6 core CPU and dual D700's in it. Geek bench has the cMP at 29000 points, and the nMP at 20000 points. The same is true for Luxmark, I didn't write the scores down but the cMP beats it. So I figured as best the nMP would be a close second.

Both of these machines have just 16GB of RAM, the same OS, same software version and photo set to work with.

My cMP was able to build a Dense Point Cloud in 483 seconds, which seemed fast enough to me. But doing the same workflow in the nMP it only took 291 seconds, despite being about 33% slower in almost all benchmarks. This software uses openCL and all graphics cards are utilized.

This was a small test project too, using a larger data set of hundreds of photos the cMP took over 80 hours to complete. I may run a test and see how long the nMP takes on that same data set.

 

[mcnallym]

You will probably find that as the Software you are using is OpenCL and seems to use the GPU's for this then the software has been optimised/rewritten/updated to make use of the nMP.

This is the potential for the nMP but it does require that software is written to make use of the Hardware available. As software is updated then may start to find that the nMP starts to look a better option to the cMP.

Would be interesting to see what the cMP would be like with a pair of 7970's or GTX970's.

 

[Macinsquatch]

Well threaded software that takes advantage of openCL of the installed GPU isn't really machine specific, just component specific. Most likely a function of faster bus speeds, faster RAM, and GPU's with a bit of openCL horsepower in the nMP. Wish I had more GPU's to test in the cMP.

There are plenty of instances where my cMP does to opposite and kills my nMP too.

 

[crjackson2134]

Which GPU do you have your monitor attached to on the cMP?

 

[Synchro3]

Do you have installed a PCI Express SSD in the cMP? Only with a PCIe SSD it would be a fair comparison, e.g. same i/o.

 

[IowaLynn]

Do you have installed a PCI Express SSD in the cMP? Only with a PCIe SSD it would be a fair comparison, e.g. same i/o.

+1

That 1100MB/sec or above. One of the bottlenecks that Classic and most systems: I/O and getting data to/from processor (memory is just way-station middleman)

 

[Macinsquatch]

Do you have installed a PCI Express SSD in the cMP? Only with a PCIe SSD it would be a fair comparison, e.g. same i/o.

If I were editing video I would agree, but since this is a computationally intensive task it really makes no difference. Like I said, the cMP kills my nMP in all synthetic benchmarks, a PCI SSD would further that lead.

Edit: One thing that will really improve performance on my cMP will be to get a triple channel memory configuration. With only 2 8GB sticks my RAM speeds are about half of what they should be. I don't plan on getting rid of this beast so more RAM is in the future.

----------

Which GPU do you have your monitor attached to on the cMP?

I have it hooked up to the 970.

 

[sirio76]

If I were editing video I would agree, but since this is a computationally intensive task it really makes no difference. Like I said, the cMP kills my nMP in all synthetic benchmarks, a PCI SSD would further that lead.

Edit: One thing that will really improve performance on my cMP will be to get a triple channel memory configuration. With only 2 8GB sticks my RAM speeds are about half of what they should be. I don't plan on getting rid of this beast so more RAM is in the future.

----------



I have it hooked up to the 970.

Adding RAM will not improve significantly the speed of your system.

 

[AidenShaw]

Adding RAM will not improve significantly the speed of your system.

It will significantly improve it by a small amount.

 

[sirio76]

???

 

[AidenShaw]

???

In statistics (which is the topic here), "significantly" does not mean "large".

A definition at dictionary.com is

Statistics. of or relating to observations that are unlikely to occur by chance and that therefore indicate a systematic cause.

It will be slightly faster with more channels. It is unlikely on actual workflows that it will be noticeably faster.

 

[ActionableMango]

There's no better benchmark for the task at hand than actually doing the task!

 

[Macinsquatch]

So I did a test with a larger data set, 344 photos to be exact this time. No changes to either machine as far as RAM, OS, or software used.

cMP was able to generate the dense point cloud in 73,923 seconds.
nMp did the same workload in 43,474 seconds.
Of note is the nMP was much more usable while working, the cMP became quite sluggish and had trouble keeping up with typing. I don't have any measurements of this but it was noticeable for the sort durations I had to use both machines.

In the end this is a fairly substantial difference, although only representative of this workload. It does make it much easier to justify the purchase of the new machine.

 

[linuxcooldude]

Yes, benchmarks are only for very basic comparisons. I have not really seen yet any good benchmarks designed for workstations. Its usually ones designed for gaming with the occasional app that actually designed for these kinds of computers. The best measure of a workstation is the actual workload itself used in day to day operations.

Even the cMP can only do so much even if upgraded and is getting long in the tooth. Great comparison based on a real world scenario.

 

[crjackson2134]

Would be interesting to see what the cMP would be like with a pair of 7970's.

Agreed, since the nMP has 2x D700s pumping the OpenCL, the cMP would need 2x HD7970s running at 5GTs to make a fair comparison on a givin app.

 

[Macinsquatch]

Agreed, since the nMP has 2x D700s pumping the OpenCL, the cMP would need 2x HD7970s running at 5GTs to make a fair comparison on a givin app.

And have one processor removed.....the nMP is only a 6 core. I'm not trying to do an apples (haha) to apples comparison.

I guess the point is that using the "standard" benchmarks that we all use as the proverbial dick measuring sticks for these machines had my cMP significantly faster. Faster in openCL luxmark scores, faster in geekbench, not by a little either. That is why I'm so surprised that my nMP has such a wide lead on my workloads.

 

[crjackson2134]

And have one processor removed.....the nMP is only a 6 core. I'm not trying to do an apples (haha) to apples comparison.

Yup, very true...

 

[ActionableMango]

I guess the point is that using the "standard" benchmarks that we all use as the proverbial dick measuring sticks for these machines had my cMP significantly faster. Faster in openCL luxmark scores, faster in geekbench, not by a little either. That is why I'm so surprised that my nMP has such a wide lead on my workloads.

I'm surprised that the OpenCL benchmark had the cMP faster if you have better OpenCL resources on the nMP. I wonder if the benchmark application is unable to take advantage of multiple GPUs.

 

[faefae]

If I were editing video I would agree, but since this is a computationally intensive task it really makes no difference. Like I said, the cMP kills my nMP in all synthetic benchmarks, a PCI SSD would further that lead.

Edit: One thing that will really improve performance on my cMP will be to get a triple channel memory configuration. With only 2 8GB sticks my RAM speeds are about half of what they should be. I don't plan on getting rid of this beast so more RAM is in the future.

----------



I have it hooked up to the 970.

Dude you are wrong. Read about the IO wait state. Even cpu oriented workloads benefit from ssds. In those situations it is about the benefit of lower storage latency, not necessarily bandwidth.

 

[fhturner]

Only realized down here at the end that most of this conversation took place almost 2 years ago. But my thought goes to the storage as well...especially the experience of choppiness w/ typing, which sounds to me like it could be virtual memory thrashing. Appears that you're working (or WERE working) w/ some pretty large datasets, so I'd be interested to see the memory usage during the runs, since you mention both machines have "only" 16GB (which may indeed be small for the size of your dataset).

If both machines run out of memory and have to resort to VM, and if the cMP is using an HD (or not a very good SSD), the nMP then commands a MONSTROUS advantage at this point, which could explain the surprising results. That x4 PCIe SSD in the nMP would compensate for the virtual memory paging slowdown orders of magnitude more effectively than an HD. And that's not even down to raw transfer rate— it's not so much the *speed* (e.g. ~1500MB/s) that makes SSDs so effective, it's the *quickness* (e.g. ~100,000 IOPS). Hard drives just can't jump around to different chunks of data anywhere near as fast.