Welcome to our community

Be a part of something great, join today!

Log in
  • Hey all, just changed over the backend after 15 years I figured time to give it a bit of an update, its probably gonna be a bit weird for most of you and i am sure there is a few bugs to work out but it should kinda work the same as before... hopefully :)

5K iMac or New Mac Pro?

jussi rovanpera said:
Regarding the Mac Pro gpus: Resolve can only use memory of 1 gpu card, so 2 x 6gb is 6gb in resolve. It can use the processing power of the 2 cards.
Click to expand...

This is not true. Resolve uses the RAM on each GPU, after all how could the GPUs perform their computations if not using their onboard memory? What Resolve does, however, is restrict the amount of RAM used to that of the amount of the smallest GPU RAM size in the system. So if you have two GPUs in the system and one has 4GB, the other has 6GB, then it will only allow for 4GB used on each.

2 X 6GB in Resolve is not 12GB, it's 2 X 6GB -- you have two separate GPUs sharing the load, each with their own respective 6GB of workspace in which to operate.

A GPU arrangement with one 4GB and one 6GB is the same as 2 X 4GB.

In many cases, lots of duplicate data needs to be loaded to each GPU in order for the GPUs to do the necessary processing. Or you have data for subsequent frames or frame segments loaded to each GPU. If you're working with 8K images that don't fit within a 4GB frame buffer, having multiple 4GB GPUs isn't going to fix this.
 
Theoretical vs Actual Bandwith

Theoretical vs Actual Bandwith

Blair S. Paulsen said:
Correct me if I'm wrong but isn't 4x PCIe3 equivalent to 16x PCIe2? Not saying there's no bottleneck there with the latest GPUs, just that saturating 4 lanes of PCIe3 is a pretty serious amount of bandwidth. I concur with Eric's preference for real PCIe3 slots over any external bus attachment, but how often will that be the critical bottleneck? Looking forward to seeing tests of eGPU vs slotted GPU with Pascal cards on typical media creation tasks.

Cheers - #19
Click to expand...

So, based on the think tank discussion here -- the reasonable expectation will be that eGPU solutions can add wheels to situations that have none, but the classic workstation will still be required for heavy lifting. I can live with that. It can beef up/extend the outer ring of Apple computing tools - Air, MBPs, Minis, actually anything non trashcan, and still get a decent/flexible computing experience.
 
Thunderbolt 3 specifically allows for eGPUs in the spec. Whereas with TB2 and TB1, it was just do-able in Windows as part of PCIe expansion. So this should help on the Mac platform to get eGPUs on the market.

TB3 requires new cables due to the new channels/ increased bandwidth. Not to mention it's based on USB-C connectors instead of mini-DisplayPort.

The 40Gbps bandwidth is roughly equivalent to PCIe v1 X16. We also have to consider the overhead and what is lost in the PCIe distribution -- we're really only going to get about 35Gbps of PCIe bandwidth. 40Gbps is still available in full for video stream pass-thru, just as 20Gbps full was available for video streams on TB2, as the video stream bandwidth is not specifically limited by PCIe bandwidth.

This may still be plenty of bandwidth for a lot of computational eGPUs. For raw GPU performance, however, it's lacking -- only providing about 22% of the effective bandwidth offered by current PCIe v3 X16 slots.
 
Eric Bowen said:
Keep in mind TB3 is just a PCI-E 3.0 4X slot. That means external GPU's would have a very limited highway to handle all the data they normally handle with current GPU acceleration models. I would not suggest planning on that as a viable alternative to actual PCI-E slots.
Click to expand...
Ah, that's good to know. So much for that! I really wish that Apple would just make (say) a 4RU rackmount Mac with PCI-E slots. They have it in their pointy heads that they have to make fashion statement machines instead, and that kills a lot of pro users in terms of expandability. I'm still bitter that they went from 6 slots to 3 slots in the old desktop line.

Note that there are quite a few major Resolve shops using 8 Titan X GPUs, and those are screaming fast. But as far as I know, that configuration only works under Linux right now. I tend to doubt that Apple would do the work necessary to allow OSX to talk to anywhere near this many GPUs.
 
You would think Apple would want to tout the fact that they have the best machines for professionals, even if just a small percentage of their market will buy them.
We basically want Apple branded HPZ800s.
 
Jeff Kilgroe said:
This is not true. Resolve uses the RAM on each GPU, after all how could the GPUs perform their computations if not using their onboard memory? What Resolve does, however, is restrict the amount of RAM used to that of the amount of the smallest GPU RAM size in the system. So if you have two GPUs in the system and one has 4GB, the other has 6GB, then it will only allow for 4GB used on each.

2 X 6GB in Resolve is not 12GB, it's 2 X 6GB -- you have two separate GPUs sharing the load, each with their own respective 6GB of workspace in which to operate.

A GPU arrangement with one 4GB and one 6GB is the same as 2 X 4GB.

In many cases, lots of duplicate data needs to be loaded to each GPU in order for the GPUs to do the necessary processing. Or you have data for subsequent frames or frame segments loaded to each GPU. If you're working with 8K images that don't fit within a 4GB frame buffer, having multiple 4GB GPUs isn't going to fix this.
Click to expand...

Sory, bad wording on my post. My understanding is that if you run out of 4gb GPU memory, you would also run out of memory on 2 x 4gb system, but maybe that's an oversimplification of how it works, I have not stress tested 2 gpu vs 1 gpu systems.
 
Jarek Zabczynski said:
You would think Apple would want to tout the fact that they have the best machines for professionals, even if just a small percentage of their market will buy them.
We basically want Apple branded HPZ800s.
Click to expand...

And Apple could do so much better than the HP Z8x0 workstation... There is so much design by committee and a ton of corporate must-have compromises in those systems. Totally infuriating.

jussi rovanpera said:
Sory, bad wording on my post. My understanding is that if you run out of 4gb GPU memory, you would also run out of memory on 2 x 4gb system, but maybe that's an oversimplification of how it works, I have not stress tested 2 gpu vs 1 gpu systems.
Click to expand...

An oversimplification, sure. But you're right. GPU memory sizes are restricted to match the smallest in a set and they are used locally on their respective GPUs. If you run out of GPU memory with a 4GB GPU, then adding another GPU with 4GB does not help as it is also only able to provide 4GB for its own use and would fail the same way when presented with the same data. If you were to add a GPU with 6 or 8 GB, that would also not help as that GPU would be limited to 4GB as well so that conditions across all GPUs in the system can be held consistent. Only way to solve it is to add a GPU with enough RAM and disable or replace the other lesser GPU.
 
Is the lowest common VRAM issue only a Resolve problem? I used an old nVidia 120 card to get a boot screen for installing the web drivers for the 4GB GeForce 980 - then just left it in slot 4 of my 5,1 MP tower. I was actually considering swapping it out for an M.2 SSD on PCIe card, but hadn't gotten around to it. Should I pull it? I don't have Resolve on that machine right now, but would want the option to install it and get decent performance.

While we're on the subject, is there a good rule of thumb for how many GBs of VRAM one needs as a ratio to frame size? Does it scale linearly?

Cheers - #19
 
Blair S. Paulsen said:
Is the lowest common VRAM issue only a Resolve problem? I used an old nVidia 120 card to get a boot screen for installing the web drivers for the 4GB GeForce 980 - then just left it in slot 4 of my 5,1 MP tower. I was actually considering swapping it out for an M.2 SSD on PCIe card, but hadn't gotten around to it. Should I pull it? I don't have Resolve on that machine right now, but would want the option to install it and get decent performance.

While we're on the subject, is there a good rule of thumb for how many GBs of VRAM one needs as a ratio to frame size? Does it scale linearly?

Cheers - #19
Click to expand...
GT120 is only used to drive GUI, so it really makes no difference, if you have it or not. Just make sure, if you continue to use GT120, to un-click the "use display GPU for compute" in Resolve preferences. Frankly, right now, 120 just occupies the slot without adding anything. I'd pull it out.
I'm pretty sure the rule of thumb, your VRAM should be at least equal to the rendered frame size. 4GB for 4K. But with use of NR, even that amount of VRAM is not a guarantee, that Resolve will not puke...
 
Yeah, only matters for compute GPUs and you would never use the GT120 for that. It's not specifically a Resolve issue, but there were such limitations in both CUDA and OpenCL for a long time. CUDA now allows for varying the amounts of RAM and operations per GPU, but software hasn't really caught up to that. For now, when you install a new GPU that is a lot more powerful, not just in terms of the amount of RAM, it's going to be held back by any other compute GPUs you're trying to use. For compute GPUs it's best to keep them all the same or use the most powerful while disabling the others.

As for the GT120, I'd get rid of that thing unless you need it for something specific. Even if you start using Resolve on that system again, you won't be needing the crappy old GUI GPU. You can use one of your other GPUs as both display and compute and there is very little performance hit for that. However, it does affect RAM usage. So if you're pushing the limits on RAM, then you may want to consider a dedicated GUI card... Personally I don't use one on my primary Resolve box -- I'm using 3X 980TI cards and having great results. Dual Titan X in my X99 system also screams, but we fall short on cores with that system.
 
Jeff Kilgroe said:
Personally I don't use one on my primary Resolve box -- I'm using 3X 980TI cards and having great results. Dual Titan X in my X99 system also screams, but we fall short on cores with that system.
Click to expand...
Hey Jeff.
What computer rig do you use to run 3 GPUs?
I'm looking to migrate to a new system from Z820. There are not enough slots and not enough power connectors. I'd rather not go the expander route. I'm thinking of moving to Supermicro SYS-7048GR-TR 4U Server with X10DRG-Q Motherboard.
https://www.google.com/search?clien...r+with+X10DRG-Q+Motherboard&ie=UTF-8&oe=UTF-8
What do you think?
 
jake blackstone said:
Hey Jeff.
What computer rig do you use to run 3 GPUs?
I'm looking to migrate to a new system from Z820. There are not enough slots and not enough power connectors. I'd rather not go the expander route. I'm thinking of moving to Supermicro SYS-7048GR-TR 4U Server with X10DRG-Q Motherboard.
https://www.google.com/search?clien...r+with+X10DRG-Q+Motherboard&ie=UTF-8&oe=UTF-8

What do you think?
Click to expand...

Those motherboards are awesome for GPU based workflows. I've been using that series for years with 3 GPU and a Rocket-X, with 3 slots to spare for other stuff.
 
jake blackstone said:
Hey Jeff.
What computer rig do you use to run 3 GPUs?
I'm looking to migrate to a new system from Z820. There are not enough slots and not enough power connectors. I'd rather not go the expander route. I'm thinking of moving to Supermicro SYS-7048GR-TR 4U Server with X10DRG-Q Motherboard.
https://www.google.com/search?clien...r+with+X10DRG-Q+Motherboard&ie=UTF-8&oe=UTF-8
What do you think?
Click to expand...

I'm running a SuperMicro X9 series motherboard (X9DAI, I think, but I'd have to check). With dual 10-core E5 v2 Xeons, 256GB RAM and 3 X GTX980TI GPUs. It's still holding up well, even though it's a few years old now. I had GTX780 GPUs in there initially, but swapped those out with 980TI cards. Originally the plan was go go dual Titan X, which I did in a couple other systems, but I decided on the 3x980TI arrangement mostly because I felt it served the system well and NewEgg had a deal on the 980TIs. I'm planning to build something to replace this system either later this year or early next year. Then I can capitalize on DDR4 memory, newer CPUs with more cores and newer GPUs. But I'll have to see. It's been increasingly difficult over the last few years to get proper ROI out of these dual Xeon monster systems.

As for the X10DRG0-Q motherboard. I personally have not used it. But it would be one that I would definitely consider for a new build. I've been using SuperMicro boards for a long time and have always had great results with them. The only other dual Xeon boards I would consider right now are ASUS. ARIMA makes a few good ones as well, but it seems they keep the best models for integrators only -- BOXX uses them and so do a few other OEMs.
 
Jeff Kilgroe said:
I'm running a SuperMicro X9 series motherboard (X9DAI, I think, but I'd have to check). With dual 10-core E5 v2 Xeons, 256GB RAM and 3 X GTX980TI GPUs. It's still holding up well, even though it's a few years old now. I had GTX780 GPUs in there initially, but swapped those out with 980TI cards. Originally the plan was go go dual Titan X, which I did in a couple other systems, but I decided on the 3x980TI arrangement mostly because I felt it served the system well and NewEgg had a deal on the 980TIs. I'm planning to build something to replace this system either later this year or early next year. Then I can capitalize on DDR4 memory, newer CPUs with more cores and newer GPUs. But I'll have to see. It's been increasingly difficult over the last few years to get proper ROI out of these dual Xeon monster systems.

As for the X10DRG0-Q motherboard. I personally have not used it. But it would be one that I would definitely consider for a new build. I've been using SuperMicro boards for a long time and have always had great results with them. The only other dual Xeon boards I would consider right now are ASUS. ARIMA makes a few good ones as well, but it seems they keep the best models for integrators only -- BOXX uses them and so do a few other OEMs.
Click to expand...
I like to stay a generation behind on CPUs and everything that goes with it. I find it so much better in terms of ROI, than the bleeding edge. I just scored two E5-2660V2 10 core for $500. I'm happy to go with Supermicro. I'm just concerned with the noise level. I don't have a machine room, so that is important to me. I was told, that whisper quite type Supermicro enclosure type is the way to go. Do you know, if that is just the type of fans used or it is actually a different enclosure design? And if that is the case, can other type of Supermicro enclosure can be retrofitted with those?
 
You must log in or register to reply here.
Back
Top