Update V - 08/01/2013 - Knowledge = Power!
Bet you didn't see this one coming! Yes it's a minor one this time, it's going to be a performance analysis (you can skip this if you're not interested, I don't blame you ) and attempt at showcasing the system while in operational mode (and hopefully fixing the color of the coolant). So, let's go !
Gaming vs Workstation vs Server
I already touched this subject a bit in the first update but I'm building a workstation here, not really a gaming rig or server. The difference is that a gaming pc specializes more in high clock frequency and therefore low latency parts.
Games are coded in a way that makes it hard to scale across multiple threads. Also, responsiveness is needed and communication between nodes introduces latency and can kill gaming experience. I'm pretty sure that at this point in time 4-8 cores will suffice and more than 12 threads are not even used.
A work server is on the other side of the spectrum, high reliability, generally even more cores, bandwidth and RAM. The applications scale endlessly and are very much parallel which makes it attractive to run them in nodes. They have to do mission critical work more often than not and run higher density RAM which rises the need for error correction ECC etc.
A workstation is usually in between. The 3D modelling and rendering applications support much more cores than typically used in games but for responsiveness, low latency is prefered over raw bandwidth and processing power (previewing, viewport interaction).
Actually, after running Pernod for a couple of days I noticed that the suite of software I used, was not really that great to show where multi processor systems really shine. Most of them were not really that RAM dependent or scaled well over multiple processors. I did included a gaming test out of curiosity though.
Test Setup is as in my sig for Pernod and Nardis respectively:
2x Xeon E5-2665
16/32 GB | dual/quad | 1600MHz Corsair Dominator Platinum
Zotac GTX 470 | FW 326
Core i7 2600K
4 GB | dual | 1866MHz Corsair Vengeance
Zotac GTX 470 | FW 313
First of all we have CineBench. This is what I'm going to be using most of the time since it is based on the stock render engine of Cinema4D also by Maxon
This bench shows a healthy increase vs Nardis. Clocks and cores scale nearly linear. A perfect world really. While working on more complex scenes and custom render engines things might be different though. Also, the "you're gonna have to change that" crowd was right. Quad channel clearly trumps dual channel for multiple processor systems, depending on the application of course. This is probably worst case scenario for Nardis in this analysis.
Then there's GeekBench. Things start to look a bit worse for Pernod. This benchmark focuses on single and mult core performance. The former in which Nardis wins handily and the latter in which Pernod shows about the same pattern as in CineBench. The problem here is that I used the free version which is limited to 32-bit and limits the capabilities of Pernod.
Integer operations are used for OS mostly, compression/decompression and encrypting (mining), floating point in most science applications, games and render applications. Memory is limited to single core anakysis in this bench. This paints a somewhat incorrect image. Also, this bench mark isn't optimized for NUMA (Non Uniform Memory Acess, basically gives both CPUs the ability to access all RAM present in the system) architectures like the one present in Pernod so it really functions like a single core 2.5GHz Sandy subsystem...
Below a list of more detailed results. The chart above summarizes this though. I lost the data set for dual channel but that would be nearly invisible either way.
I doubt I have to say much about this gaming benchmark. I was just testing how well Pernod would keep up with the 5.1GHz i7 in Nardis. Same incapable GPU used here. Not much variables for gaming performance.
Like I suspected, 3dmark doesn't see a lot of cores. I couldn't quite find how many cores were loaded since it was spread evenly across 16 threads, but the load was no more than about 30-40% per core. This results in a Physics score about equal to a 5GHz core i5 and graphics performance nearly identical to Nardis. So you can game on this system. Was a bit of a surprise to me.
Then we have the next big surprise, power consumption. Aside from sporting about 2x the performance overall on the CPU subsystem, Pernod is actually more efficient than Nardis. By a little bit only though. This is of course mostly because the 2600K was at 5.1GHz @1.47V and the E5-2665s at 2.5GHz @1.01V for CPU1 and 0.97V for CPU2. I could further reduce this probably since I can't really overclock anyways...
And as a summary we have this chart
The memory bandwidth of the entire system was measured by the STREAM app made by some CS people over at the University of Virginia. A dual Xeon C602 system has a theoretical max of 100GB/s due to having two quad channel capable CPUs in a system with 1600MHz RAM (any higher is unstable, blame SNB-EP IMC). a dual channel 1866MHz system tops out at approx. ~ 29GB/s. Nardis got 23GB/s while I got 74GB/s out of Pernod.
So there you have it. This was really intended for the people who were asking about the use case of such a system. I might do some benchmarks of software aimed more at multi processor systems and bandwidth but I'm pretty content with these results
Correcting White Balance And Night View!
So nvidiaftw gave me some advice on how to capture the true color of coolant since it was bugging me, I gave it a try, I hope I did a good job this time
Sorry for the bad image quality, but this is only to show you the color allignment.
And lights off!
Well that was it.
I don't know when the next update will be. I'll have to search around for a proper GPU but that will take time so yeah ...
Thanks for reading!
Edited by TheBlademaster01 - 9/3/13 at 3:06pm