Enter the HP Z820

[Jeff Kilgroe]

The Z820 with dual 3.1GHz CPUs (16 cores total) can play back 5K in real-time. However, that's all it can muster when working at such resolution. The RED Rocket takes a huge load off the CPUs. The wavelet decompression is the resource drain when it comes to processing R3D files. It's not the debayer process, that's relatively simple and something GPUs can accelerate and often do in some apps, but RED has left their debayer processing internal to their own API's.

I'm hopeful that we'll see more hardware acceleration options in the near future with the Xeon Phi and upcoming Tesla compute options as they should be better suited for wavelet decompression.

FWIW, I'm running an EVGA GTX690 in my Z820 now. It's giving me the best overall performance per dollar per slot of anything I've tried. Premiere likes it, AE likes it, Resolve likes it. These new nVidia Kepler cards are similar in overall CUDA performance compared to the previous Fermi architecture, now that drivers are getting better optimized. If you need double precision or better FP32 or PF64 computations, the Fermi-based CUDA cards are still a better choice. However all the apps I'm using, which include those I just mentioned above, are all single-precision and scream along just fine.

nVidia will be releasing new Tesla cards based on Kepler over the coming months and the Tesla K20 due early next year will support higher precision. We should see updated Quadro cards after that.

After flopping back and forth between the GTX690 and various other GPU combinations, this is what I've settled on. The Quadro 6000 + Tesla C2075 (or GTX570 companion) was also a compelling dual GPU solution for Resolve, but after seeing the GTX690 in action here and it only taking one slot, I couldn't pass that up. The Quadro does hold an edge for complex OpenGL visuals and massive computational sets as it has more onboard RAM with ECC. But I'm not really needing any of that at the moment with this system. Precision on edge calculations and anti-aliasing or general pixel-precision within OpenGL, as well as performance with many legacy OpenGL operations like glReadMemory and glReadPixels are faster on the Quadro cards. If I were spending most of my time in Pro/E or SolidWorks or even XSI, I would stick with the Quadro. But this is primarily an Adobe CS6, Redcine and Resolve workstation. GTX screams.

[Andrew Madu]

Jeff,

Are you running your HP Z820 workstation with un-buffered DIMMs (UDIMM), registered DIMMS (RDIMM) or load-reduced DIMMS 1333 MHz (LRDIMM)?

From my understanding of un-buffered v buffered/Fully-buffered memory, the LRDIMM would be the way to go for graphic intensive work, correct?

Many thanks in advance,

[Jeff Kilgroe]

1600MHz RDIMMs -- I believe that's all HP offers with the top dual CPU configurations. You must use RDIMMs to get to many of the larger memory configurations over 256GB and they're recommended by Intel for anything 128GB and larger on the Romley platform. I have 256GB in my Z820. 128GB wasn't enough... seriously.


This thread is old, I have gone back to the Quadro 6000 as it delivered much more stable performance than the GTX690. Only app that suffered in speed was Resolve, but by little. I plan to upgrade to a newer Quadro and Tesla (or Intel MIC) combination sometime in the coming months.

[Danai Chutinaton]

Jeff,

I've just started using the GTX 690 because of the ability to use dual GPU's in one slot. I have an Asus P9X79 Pro motherboard.

So far its been doing great in Resolve. Haven't had time to test it much yet with Premiere and AE. Could you please tell me where you were seeing instability problems? I'm running the latest driver recommended in the latest Resolve configuration guide (can't remember the version number off the top of my head).

[Jeff Kilgroe]

The 690 handled fine in Resolve. When I said Resolve suffered in speed, I meant that it suffered when I replaced the 690 with the Quadro. The 690 or what amounts to dual 680's running at 95% of default clock speed, is faster in Resolve.

The Quadro has greater stability and also much better performance for pro OpenGL applications like Modo and XSI. Premiere and Resolve like the 690 just fine. AE seems to still have some instabilities with the 690 and even though it sees dual GPUs and uses them, the OpenGL performance on the Quadro 6000 wins out here. Although, that's just what I can tell from some simple tests, I don't use the 3D OpenGL stuff in AE much, if ever.

[Andrew Madu]

1600MHz RDIMMs -- I believe that's all HP offers with the top dual CPU configurations. You must use RDIMMs to get to many of the larger memory configurations over 256GB

Load-Reduced DIMMS (1333 MHz) are offered as a configuration option for both Xeon E5-2690 and Xeon E5-2687 CPU's, starting at 256GB.

Still can't figure out why the Nvidia Tesla C2075 is NOT supported as an option with dual E5-2687W cpu’s, but is supported when running dual E5-2690 cpu's!

[Danai Chutinaton]

Thanks Jeff!

[Jeff Kilgroe]

Load-Reduced DIMMS (1333 MHz) are offered as a configuration option for both Xeon E5-2690 and Xeon E5-2687 CPU's, starting at 256GB.

Still can't figure out why the Nvidia Tesla C2075 is NOT supported as an option with dual E5-2687W cpu’s, but is supported when running dual E5-2690 cpu's!

But why would you want to run 1333MHz RAM with those CPUs??? Anyway, I'm looking at the Z820 configure page right now at hp.com and I only have one word to describe what I'm seeing... curious. All the 32GB module options are 1333MHz LR. I would surmise that HP has yet to release or validate 32GB 1600MHz RDIMMs for the system. Something that was supposed to happen by now. Or perhaps it didn't work out as planned. I suppose I could ask the guys at HP. ;)

Back when I ordered, there were no 32GB options at all. Actually 1600MHz RAM was in serious short supply then too.

As for the Tesla C2075, I'm not sure why that option gets the red warning with the 2687W's selected. I will definitely ask the guys at HP about that. It was most certainly a config option when I ordered, as I ordered mine that way! Quadro 6000 + C2075. Then again, that was nearly 8 months ago, maybe they found a reason to not support that configuration?

I tested several different configurations with GPU cards ranging from dual GTX580's to the single GTX690. I currently have the system configured with only the Quadro 6000. Runs fine as it is, but will probably upgrade to the newer K6000X when it's released in the coming months. I sold the Tesla card. When it all comes down to it, sure it worked fine, but I found very little value to it for the apps I run -- other than Resolve -- and I was a PCIe slot short with it installed. Dual 580's took up the same space and consumed more power than the Quadro 6000 + C2075. Of course the GTX cards are not supported either. I guess what confuses me is that the Tesla card was an option with the 2687W's in the past and I have the order paperwork to prove it. I'll have to see what they say.

It's funny that back in July when I posted to this thread, I was having great experiences with the GTX690 in Resolve and Premiere. I had to remove the card shortly after as other issues were popping up in relation to the driver sets needed to run the card. Most all were known bugs, but I pulled the card and have not gone back, simply because the Quadro runs just fine. Resolve was faster having the dual GPU in a single slot, but that's really the only software that benefited from it. OpenGL performance within Pro/E, Modo and other apps I run is superior on the Quadro. Complex node arrangements in Resolve did seem to bog down the 690 more so than I see with the Quadro and I blame the smaller RAM capacity on the 690. Then again, it could have been the system and cacheing of R3D frames as I upgraded to 256GB RAM about the same time I shifted to Resolve 9 and back to the Quadro card.

[Andrew Madu]

But why would you want to run 1333MHz RAM with those CPUs??? Anyway, I'm looking at the Z820 configure page right now at hp.com and I only have one word to describe what I'm seeing... curious. All the 32GB module options are 1333MHz LR. I would surmise that HP has yet to release or validate 32GB 1600MHz RDIMMs for the system. Something that was supposed to happen by now. Or perhaps it didn't work out as planned. I suppose I could ask the guys at HP. ;) .

In regards to your opening question you may find the following video of interest: LDRIMM Controller from Inphi Increases DDR3 DIMM Capacity - IDF 2012, which in conjunction with the following info What is LR-DIMM , LRDIMM Memory? may shed some light on why 32GB 1600MHz RDIMMs have yet to appear on HPs options.

I would be very interested to hear the rational from HP in regards to some of the product combinations they no longer support.

Many thanks in advance,

[Jeff Kilgroe]

Andrew, those links, along with other info I've dug up so far pretty much shows what's going on. It doesn't seem to me that anyone has released 1600MHz 32GB RDIMMs yet. Not even sure if a viable memory controller exists to make those modules possible. Technically the memory host within the XEON E5 platform can support that RAM if it existed. Micron, Samsung and Hynix have 1600MHz 32GB RDIMMs on their product schedule, but not available. Micron lists them as in the sampling stage. It seems they're all using the new 8Gb memory chips, so doubling the density from what's on the market now in most modules.

I'm typing up an email to send to my guy at HP with a few questions, we'll see what he says. But I think I know what the answer is going to be here. The technology or necessary components to make these modules just doesn't exist yet. Look on the bright side, by going with the LRDIMMs, they draw less power and the system is already a massive power-hog as it is. :)

Mostly I'm curious about the Tesla option with the 2687W CPUs. Not sure what's going on there. It seems some other options have changed since I ordered as well.