Graphics cards

Support Technical Support
1

At the risk of being flamed for asking “obvious” questions or questions already asked (which I did search and found some similar but they usually evaded a clear answer), I have a couple questions.

I understand that scenarios of use will change the results, a line of best fit can be found across all scenarios and that is what I am looking for.

Background: Motherboard is a W97-WS with 16GB RAM. It was built to run Solid Edge CAD. It initially had a workstation PNY K2000 vid card but currently has the K2000 and two (2) K2200 with room for a fourth. Also, there is no SLI. Repeat…no SLI. The workstation cards are not capable of it.

The first question is desktop vs work station. In researching, obviously, the more CUDA cores you have, the better. But I also found suggestions that for more professional projects, the use of work station cards may be something to consider.

So in general, is there any benefit to having a workstation card (for example the 4GB K2200 with 680 core) if for the same cost one can get a desktop card (say 8GB) that has more cores (around 1600) available?

The other question came about while searching those similar topics on this site. And being a new user, I cannot link to those threads.

As stated, the machine currently has three video cards in it. Two of the same size and one smaller. And I have been asked to add/replace cards to improve performance. In one post on the site, an individual made a statement that seem to imply that a machine running multiple video cards that are not the same size, does not get any benefit from the larger cards or it implies that having more than one card, regardless of size does not give any benefit.

My questions are just a clarification on this. Again, this is not about SLI. The other posts I saw seemed to get caught up in it.

So there are multiple parts. The first is whether having multiple cards actually provides any benefit? The implication in the quote suggests that it does not, at least without further explanation.

Next, assuming that having multiple cards does have a benefit, if I run four cards all the same at 4GB, I don’t have 16GB available, but instead, I have four cards each capable of handing a 4GB tile? That is, it is not the sum of the cards memory, but the maximum that each card can contribute. Thus if I have only one card at 4GB and it takes 4 minutes to do a tile, adding a second card, also 4GB does not change the 4 minutes per tile but cuts the overall time in half since the tiles are split across two cards.

The other part then is if each card is a different size, you are not actually relegated to only being able to use one card, the one with the smallest memory, but that each card is limited to the memory limit of the smallest card. You’d still have four cards, but if the smallest was 1GB and the others were 4GB, you’d in essence have four cards 1GB each. So overall, you’d still get the benefit of multiple cards, but there would be a lot of unused memory.

Am I understanding everything correctly? Having multiple cards does indeed help, but the overall contribution from each card is limited to the size of the card with the smallest memory. Or have I missed the mark?

Sorry for the long winded topic, but if you don’t ask detailed questions, you rarely get detailed answers.

2

Yup, you sort of summed it up:

  • cards are as nodes in a network

  • while all running at once, file/scene is compiled to be rendered for the card with the lowest amount of vRAM

also

  • you can have as many files/scenes set as you have cards/nodes, dividing by frames/layers/scenes to suit specific goals… all cards working to full potential while still utilizing CPU (to render another, if needed)
3

workstation cards are optimized for rendering and stuff, so yes, performance will be better.

the qoute is true for non-sli. think of it like network rendering, each card needs access to its own copy for maximum performance. that way your not limited by the pcie and other components. each card can then fully operate like seperate cores, effectively doubling the speed for 2 cards, quadrupling for 4.

4

Daedalus:

So would you value a lower memory workstation card over a higher memory desktop card? A 4GB workstation card has about 680 cores while the 8GB has around 1900. Does the optimization of the workstation card make up for the lesser number of cores?

5

Burnin:

Thanks for your reply. I was pretty sure that multiple cards did help but all the posts I saw did not specifically explain it to that point. It only seemed to make sense that it was much like a RAID setup. You can have any size of drives you want, but the RAID will only be sized to the smallest drive and any drive with more than that is left with unused memory.

Thanks again for your time.

6

Unless I have missed something in recent development, workstation cards (Quadro series) get beaten in cuda rendering in everything except memory size so buying a Quadro card only makes sense if you need massive GPU memory that new cards offer. Bang for buck factor for them is way lower than for gtx cards. The fact that they are optimized for 24/7 run, double precision, opengl etc contributes nothing to Cycles rendering.

7

IMO, Quadros makes sense only with works that demand more power, longer use & higher precision in simulations or final works, ranging from prints to screens and other media, which is why is valued more.

8

dont the quadro cards have low latency stuff? i personally wouldnt buy a workstation card unless i needed a crazy amount of vram that wasnt in consumer cards.

9

Well, they have all kinds of stuff but Cycles rendering does not use that mostly. Like the double precision for example. Quadro cards have higher flops in double prec but cycles uses single precision so it is useless.

I think that to give more meaningful suggestions we need to know what exactly should the new card achieve (be used for). Is it cycles rendering, viewport acceleration, other 3D/CAD applications besides Blender that are maybe more important and so on.

As topic starter described the computer as a workstation, not a render server, I doubt that it will be continuously rendering (days on end) whuch woulf maybe justify the 24/7 workload specs of quadro cards.

10

Thanks for the discussion.

Kensonmis is correct, this computer is a CAD (Solid Edge) workstation and not a render server. When rendering, it is cycles rendering that is being done. The load should be minimum, with only a few long renders (greater than a day). In general, during the week, only minimum use is given and if a long render is to happen, it is usually done over the weekend (but not every weekend) and there is plans to do a long render over the Christmas holiday. So far, only one long render has been done a month.

Based on what has been said, I think we’ll just get some additional K2200 workstation cards. While Solid Edge can use any video card, the Quattro line is one tested to be optimized for Solid Edge. We know through experience that they do not cause problems and allow our engineers to work smoothly without delays or jerking. Thus, I have to keep one workstation card for Solid Edge.

Due to the following:

-Blender limits the available memory of every inserted memory card, regardless of size, to the size of the smallest card
-I have to keep a 4GB workstation card for use with Solid Edge (though it probably would work fine with any card)
-Blender does not seem to allow a user to pick a subset of cards, like using only 3 of 4 cards…you can pick any individual card or select all cards. No in between.
-I was looking at some desktop 8GB cards

I think that I will just end up getting some more of the K2200 workstation cards. It seems pointless to purchase cards with greater memory if half of that memory will not be available because of the smaller card or the inability to pick select subsets of cards to use with Blender that allow me to pick every card but the smallest one.

Correct me if I am wrong, but we did not see a way to pick a subset of memory cards. Anyone know a way? If there is, I can get three larger memory desktop cards and fully use them for Blender while retaining the workstation card for Solid Edge.

Again, thank you for sharing your experience.