8k r3d editing workstation.

[Barmak Armand]

Hello,
I'm working on a new workstation rig to render 8k files faster.

From what I understand Adobe pp does not utilize multicore processors that well, only the 2014 version of premier seems to do it the best. So I will be using davinci resolve pro edition.

Setup includes:
Duel Xeon e5-2669v3 processors
Duel 1080gtx graphic cards
128gb ddr4 ram
4x Samsung 850 1tb Evo raid
2*16 TB hardrive for storage
Supermicro duel socket 2011v3

My main concern is whether Davinci Resolve will utilize the cores to make render speed times faster. I know that an 8 core overclocked i7 out performes a 22 core Xeon processor Rig with the same specs. It has been shown in tests, but I'm hoping davinci resolve isn't as horribly programmed as Adobe pp when it comes to using multicore processors.

Does anyone know from experience, whether it's worth getting duel Xeon. I don't care so much about the price, it's the speed that matters.

 

[Simon Dunne]

Go single i7. Add in another GPU, like an M4000 for UI and let the 1080's (ti version with 12gb) do the GPU rendering. 64Gb ram should be fine too.

 

[Jeff Kilgroe]

Nah... Don't bother with a separate GPU for GUI use. It wastes valuable PCIe lanes and provides little, ~5%, performance advantage for apps like Resolve that can utilize separate GUI and compute GPUs and at the same time causes other issues and difficulties with apps like Premiere which default to your display GPU for acceleration.

1080 TI with 12GB... Do they truly exist yet? Nope. The Pascal Titan X 12GB is the current king of the hill, but the performance vs. the price doesn't make them worth it. Not when 40% less you can get cards like the EVGA 1080 Classified version and get 90% of the performance.

64GB is a stretch if you're doing anything serious in Resolve or working with 6K to 8K compositing. Doubly so if you have dual 12GB GPUs. Put 128GB in the system and don't short-change yourself. If you go with a workstation class motherboard that can take a Xeon CPU, a single 14 or 18 core Xeon CPU instead of the i7 is also a reasonable consideration. If you go above 10 to 12 cores definitely consider 256GB RAM. Definitely go with 128GB RAM if you go with the 10-core i7. 256GB is overkill most of the time unless you have 20 cores, but 64GB is not enough when you get to 8 cores and dual GPUs. Unfortunately we can't add intermediate amounts that would be more preferential to our chosen core count.

 

[Jeff Kilgroe]

Going back to the original post, Resolve makes much better use of more cores than Premiere does. My primary Resolve system is an older dual 10-core Xeon setup with 3 x 980TI 6GB GPUs and it still smokes my overclocked 8-core i7 with dual Titan X (prior generation) GPUs.

The down side to the newer Xeon CPUs is the higher core count chips make huge sacrifices in clock speed. They're primarily designed for virtualization and data center applications where more cores at lower speeds is desirable. The reason for dual Xeon socket workstation and server boards is so we can get the core count up while maintaining higher clock speeds. Make sure you pick Xeon CPUs that have a base clock speed (the speed at which the chip will run at with all cores firing under load, not the turbo speed) of at least 2.6GHz. 3GHz or more is definitely preferred.

For memory, you need to plan on 4GB per thread or 8GB per core plus another 8 to 16GB for overhead and that will give you the minimum you need if you actually intend to use all your cores. So 10 cores is 20 threads = 96GB, but with quad-channel memory arrangements that pushes you to installing 128GB.

 

[Wayne Morellini]

Thank you all for answering this question. I have asked this sort of question before and gotten how long is a piece of string like dubious answers instead of hard facts.

So, how is the rumored 32 core AMD Zen Naples x86 fit into this equation? They are rumored to also be contemplating combining it with a GPU. I'm not fussed, but if it can get advantage I am.

 

[Wayne Morellini]

The issue with how many cores relates to thermal envelope at speed. Unless you make chips bigger under present silicon, you com to limitations. The solution is card CPU's like Intel used to do. But you have to integrate most everything on chip with only external IO lines and intercard communication lines. I wonder how limited the AMD server solution will be.

 

[Misha Engel]

The rat-race is on again, with a long forgotten kid in town AMD.

EPYC 7551P 32c/64t 2.0..2.6 GHz on all cores 8ch DDR4 2666 128 PCIe lanes for $ 2.100 Air cooled with (Noctua D15)
Threadripper 1950x 16c/32t 3.4..4.0 on all cores 4 ch DDR4 2666 64 PCIe lanes for $ 999 Air cooled with (Noctua D15)

A bunch of fast and very high priced new Xeons from intel.

After SIGGRAPH 2017 (30 july - 3 august) we will know with what kind of GPU's AMD and NVidia will come up with.

Exciting times.

 

[Wayne Morellini]

At toms, they tried cyro cooling. Surprisingly the Intel's were getting a bit over 6ghz, which is similar to speeds gotten years back. So say ten or more years of desity increases result in similar speeds.

But what is probably happening is they went from high frequency design to lower frequency, lower power but more per cycle design. Arms used to do even more of this. The highest energy efficiency design just let's the circuit tick over at the silicon's processes natural frequency (basically for decades things have been often overclocked). To get past this also eventually requires special design to handle higher speeds.

What this all means is that it should be possible to design higher speed chips that under cyro cooling should be able to get 10ghz, 20ghz maybe possible, but past that gets into some exotic stuff or limited circuites. What this means is that, like when competitors made mobile x86 chipsets, there is possibility to make higher speed workstation processes. But like my analysis virtually predicted, the x86 s becoming an administration processor for higher performance applications. It has become pointless, except in some cases of legacy application, as Linux based app and some non x86 design can out do it.

What will such a device look like?
If we are unfortunate, it will be some sort of liquid cooling bath (used in high end computing) or cyro bath in vertical/near vertical tiles to maximise cooling flow. But I suspect pointless because so much can be done on GPU's, so a balance between needed CPU and GPU etc is needed. All of this can be put into the bath. So, a customised system could offer advantages, simply, if you can redesign everything to run at double speed this way, you instantly get work flow speed up. Today it would have been better 8k, but by the time it could be done, we maybe on 7nm and looking into much higher need than 8k. Only 4 4k cameras; n hydrogen equals 8k, but there is belief there are lots more than 4 on the professional version.

The reality is that we are dealing with legacy software that is not mature and optimised for the best gpu/whatever performance. We can get drastic savings just by a codec manufacturer's optimised subroutines available to software manufacturers to produce optimised versions that work only on that camera's footage, which improves that camera's brand.