Help: $5K i7 machines beating $20K dual-Xeons?

[Flo Eckhardt]

check out this page : http://ppbm6.com/News.html ... Harm did this benchmark for adobe you can download ... you will see all top performers including his monster are i7. Please note it´s not the most up to date site but nice to get an idea.

 

[Christopher*Seguine]

Thats old, dual e5-2687 is faster, but that "benchmark" is effected by disk speed more than anything else.

check out this page : http://ppbm6.com/News.html ... Harm did this benchmark for adobe you can download ... you will see all top performers including his monster are i7. Please note it´s not the most up to date site but nice to get an idea.

 

[Les Dittert]

Xeon memory bandwidth is hampered by ecc memory and extra wait states. Working with images involves a lot of memory thrashing.

 

[Andrei Mihailescu]

There is no such thing as the Xeons and the Core i7 CPUs being designed for different things that would explain such performance differences. There are of course slightly different design goals between the two CPU lines, and the compromises made do affect performance to a certain degree, but the basic architecture is identical, and the systems perform very close to each other when doing identical computations.

One should however pay attention to the specs of the actual CPUs and to the configuration of the system around them, since that can and will make quite a different in performance.

The E5-2690 has a base frequency of 2.9 GHz, and can turbo up to 3.3 GHz with 8, 7 or 6 cores active, to 3.5 Ghz with 5 or 4 cores active, to 3.6 GHz with 3 or 2 cores active and to its maximum of 3.8 GHz with 1 core active. The i7-3960X has a base frequency of 3.3 GHz, and can turbo up to 3.6 GHz with 6 or 5 cores active, to 3.7 GHz with 4 cores active, to 3.7 GHz with 3 cores active and to 3.9 GHz with 2 or 1 core active.

So, right off the bat, the 3960X has a frequency advantage in most situations, which translates more or less linearly in the equivalent more performance. This effect can be exacerbated by two things: if the cooling system of the Xeons is so dimensioned that they won't be able to turbo up, but for instance the 3960X can, you could for instance have a case where the Xeons are handicaped by being staying at say 3 GHz, whereas the the i7 can go to 3.6 GHz - that would give the i7 a direct advantage of about 20%, which may not be compensated by the Xeon having 2 more cores, if the application doesn't scale linearly with more cores. The same effect can happen if for instance the Xeon motherboard has a BIOS with a bug or two that prevents the Xeons from achieving the maximum Turbo Boost they can - several server/workstation manufacturers had such bugs in various BIOS revisions. So one thing you could do is look at the Turbo Boost state of the systems while rendering typical projects, and make sure that the Xeon CPUs have the playroom to scale their frequency up and down as they should.

Another usual gotcha that people ignore when comparing systems is actual memory speeds installed (and used!) - it is not unusual to have slower speed memory in the Xeon systems, because either large ECC high-speed DIMMs are too expensive, or because the motherboard/BIOS downrates them when all channels and all slots are occupied. Check this as well, and make sure that for instance you don't have DDR3-1300 MHz memory installed (or used!) in the Xeon systems, but DDR3-1600 in the Core i7 systems. There will be a bandwidth hit from ECC memory on top of that, like others said, but if the Xeons are handicaped from the start, this will make the potential peformance difference even greater.

Also, without more detailed informations on the actual configurations, it is hard to actually say if there aren't other bottlenecks that may be even more important that all of the above, i.e. one should the make sure that the storage actually is comparable (not just raw sequetial R/W bandwidth, maybe their random I/O performance is different, and rendering can be quite randomy as well, depending on the source layers).

 

[Eric Bowen]

The threading with Adobe is actually decided by workflow and Codecs. Many of the effects do not thread well at all or do not thread beyond 1 thread. Many of the input options such as key frames do not thread beyond 1 thread. When I asked Adobe about this the response was much of the input commands are to difficult to predict and allow the threads to split. The Codec threading has many variables behind it but 1 main one has to do with compression level and how complex the algorithm is. Mpeg 4 for example ie H264 or AVCHD has a very high compression level and complex algorithm which makes threading more difficult. This is due to the inherent process of long GOP compression. DCT compression though similar often has far better compression ratio and is far more efficient. This allows much better threading capability ie Pro Res and DNxHD. Keep in mind though Quicktime on Windows does not perform well ie thread well and is not 64bit yet. Quicktime will vastly change the performance of a system when playing back with effects or rendering when comparing systems. Another variable to codecs with Adobe that effect the performance are the caching models currently used for that codec. Adobe is constantly changing those with updates especially in Premiere and the GPU acceleration. That frame caching and frame buffers used will greatly impact the performance of both the cpu threading and the GPU load. This is why many show very little performance difference between the mid level Geforce cards and the High Level often times when that is simply the codecs used are not allowing the higher Geforce cards to be really pushed by what is accelerated. This is also why many see Adobe not threading beyond 12 threads. This is due to codecs used and memory management of those codecs. Both Premiere and AE thread to 48 Threads perfectly fine if the Codecs and memory management allow it.

This leads to the impact of Red. Red threads extremely well in Adobe. Testing with both single and dual 12 Core E5 Xeon configs show threading across all available threads and better performance with Red 4K than a 4930K @ 4.5GHz with Playback and Export/Render. However with that same comparison AVCHD to H264, the 4930K @ 4.5GHz is faster than the Dual Xeon 48 Thread system on export/Render. This also effects the GPU performance. Red with dual or single 12 core Xeons will push a Titan card while AVCHD with the same system wont. AE specifically will always perform better when Multiprocessing is enabled and enough ram is available for all the threads provided effects that are single or poor threaded are not used. I have a Test AE Comp that is 64 layers with various media such as MPEG2, Tiff, and PNG files. That takes 1.5 hours to export to AVI Lossless on a Dual Xeon 8 Core system. The same Comp takes 2.5 to 3 hours to export on a 4930K at 4.5GHz. What really separates the Dual Xeons from the single I7 systems is the Realtime Peak Performance capability. Dual Xeon systems are able to handle a far greater load realtime ie higher res media, effects, and layers before rendering the timeline is required for preview. This allows far greater expansion of workflow than the single I7 system. I have a Red production house client with several desktop and I7 workstation systems. The recent Dual E5 2697 V2 system purchased was able to handle the 5K media far easier which greatly expanded their workflow options in a marathon 30 Hour shoot/edit. However they were surprised the export times on the Dual Xeon were not faster than the I7 Workstation. That is where the delivery codec and 3rd party effects changed the result. This puts the Dual Xeons as Production systems but not the ideal render systems often with regards to Adobe. Keep in mind 4D and applications like that are completely different. The Render times are drastically reduced with far greater threads.

Combine all of this with the extremely Dynamic workflows/media used and the ideal system is completely specific to each client and their workflow. There is the 90% philosophy that Apple uses with their product line which covers 90% of the customer base they intend to market to. If a client is outside of that 90% then they are expected to alter their workflow to make it work or use another solution. That is good enough for much of the media content creation market. PC is the alt which offers the configuration options to meet what the 90% doesn't cover.

 

[Bruce Allen]

I'm getting some strange results here, working in After Effects and Premiere.

My $5K i7 machines (not even overclocked) are making mincemeat out of my $20K dual Xeon systems. The i7 machines are about 2x or 3x faster. And the dual-Xeon machines also have way, way more RAM. My main dual Xeon machine here has 256GB ram, plus a Titan, while the i7 machine has only 64GB ram and a GTX 680. Yet, the i7 is much faster for Adobe tasks.

What the heck is going on here? The main tasks I have been using are motion-tracking/stabilization in AE and editing/rendering in PP.

Any ideas why I am seeing this kind of performance discrepancy?

Because Adobe don't care about high-end enough users enough to write good multi-threaded code that uses many CPUs properly throughout.

They figure that more users will pay for shiny features than care about how well they are multithreaded.

Anyway, that business model appears to be working, sadly.

I get totally ignored every time I ask for stereo support in After Effects. I guess that stereo is considered too high end for them too.

It sucks, because After Effects is a great program and would be really nice if they put in the extra 10%.

Personally, I have currently stopped buying all After Effects plugins and started buying OpenFX plugins that work with Nuke and Resolve.

I'm not investing further in the Adobe platform if they aren't interested in supporting my needs.

Bruce Allen
www.boacinema.com