Loading...
> >

Is DSLI necessary?

donovancolbertdonovancolbert Posts: 1,421
in Technical Help (nuts n bolts)

Do you need multiple GPU cards connected via DSLI for iray to leverage them, or will it suffice to just have two Nvidia cards in your motherboard? 

​During render, you can check your GPU or your CPU or both... it seems that you should be able to just do the same thing with checking multiple GPU cards... it is the same idea as a server farm with GPUs that are virtually shared, right, except instead of a GPU rendering farm this would be multiple non SLI cards in one machine? 

​I mean, isn't that the basic concept of how bitcoin mining works? It isn't that the GPUs all talk to one another... it is that they all go into a virtual pool of GPU cycles that the software can utilize? 

This part isn't very well documented or discussed anywhere. 

 

Comments

  • namffuaknamffuak Posts: 4,191

    SLI is specifically NOT recommended for Iray; it will allow you to select and use multiple cards at render time. BUT each card runs independently - the entire scene must fit in the card's memory for the card to be used. A 4 GB card and a 6 GB card will both work on a scene under 4 GB; over 4 GB the 4 GB card will be dropped for the render; over 6 GB and only the CPU is available.

  • donovancolbertdonovancolbert Posts: 1,421
    namffuak said:

    SLI is specifically NOT recommended for Iray; it will allow you to select and use multiple cards at render time. BUT each card runs independently - the entire scene must fit in the card's memory for the card to be used. A 4 GB card and a 6 GB card will both work on a scene under 4 GB; over 4 GB the 4 GB card will be dropped for the render; over 6 GB and only the CPU is available.

    So, I can drop as many different cards, with as many different cuda cores, into my Motherboard as it has slots for, and the bottom line is, "least amount of memory will be the base card the scene uses,"? 

    Seriously. I want to believe this is true. My board is Crossfire capable, so two PCIe/16 slots, but not DSLI capable... 

    And I have a 1070 8GB card. So if I throw a 1080 11GB card in another slot... 

    The 1070 memory is the base that I'll have available, but I'll get all of the 1080 cuda cores working for me, wihtout DSLI? 

     

  • mjc1016mjc1016 Posts: 15,001
    donovancolbert said:
    namffuak said:

    SLI is specifically NOT recommended for Iray; it will allow you to select and use multiple cards at render time. BUT each card runs independently - the entire scene must fit in the card's memory for the card to be used. A 4 GB card and a 6 GB card will both work on a scene under 4 GB; over 4 GB the 4 GB card will be dropped for the render; over 6 GB and only the CPU is available.

    So, I can drop as many different cards, with as many different cuda cores, into my Motherboard as it has slots for, and the bottom line is, "least amount of memory will be the base card the scene uses,"? 

    Seriously. I want to believe this is true. My board is Crossfire capable, so two PCIe/16 slots, but not DSLI capable... 

    And I have a 1070 8GB card. So if I throw a 1080 11GB card in another slot... 

    The 1070 memory is the base that I'll have available, but I'll get all of the 1080 cuda cores working for me, wihtout DSLI? 

     

    If you have an 11 GB card and an 8 GB and a 4 GB card, as long as the scene is less than 4 GB then all 3 cards will be used.  If the scene is above 4 but less than 8, then those two will be used.  Greater than 8 and only the 11 will be used....

    BUT...Windows (especially 10) will play havoc with the total available memory on ANY card that has physical hardware for connecting a screen to the card, so actual scene sizes per card will NOT be so clear to figure out.

    So, yes you can drop in as many cards as you have slots for and they will be used as memory/scene size permits. 

  • JamesJABJamesJAB Posts: 1,760
    donovancolbert said:
    namffuak said:

    SLI is specifically NOT recommended for Iray; it will allow you to select and use multiple cards at render time. BUT each card runs independently - the entire scene must fit in the card's memory for the card to be used. A 4 GB card and a 6 GB card will both work on a scene under 4 GB; over 4 GB the 4 GB card will be dropped for the render; over 6 GB and only the CPU is available.

    So, I can drop as many different cards, with as many different cuda cores, into my Motherboard as it has slots for, and the bottom line is, "least amount of memory will be the base card the scene uses,"? 

    Seriously. I want to believe this is true. My board is Crossfire capable, so two PCIe/16 slots, but not DSLI capable... 

    And I have a 1070 8GB card. So if I throw a 1080 11GB card in another slot... 

    The 1070 memory is the base that I'll have available, but I'll get all of the 1080 cuda cores working for me, wihtout DSLI? 

     

    If a card is included in the reder job, all of it's Cuda cores will be used, regradless of how many cores the other cards have.

    Keeping that in mind.....

    Let's say that you have 3 cards (1x GTX 1080 ti 11GB, 1x GTX 1080 8GB, and one GT 630 4GB) and your render job will fit on the 4GB card.
    Including the GT 630 in the render could slow your render speed the same as if you included a slow dual core CPU in the mix.
    This happens because of the huge performance gap between the fastest and slowest device:
    GTX 1080 ti (3584 Cuda cores)
    GTX 1080 (2560 Cuda cores)
    GT 630 (96 Cuda cores)

Sign In or Register to comment.