I have 670 GTX with 2Gb of VRAM. Thinking of upgrading to 1060 with 6Gb of VRAM. Is it worth it or should I save up for 1070 with 8Gb of VRAM ?
Thanks
P.S. For sculpting and Cycles and whatever else is on GPU in Blender.
When u need the 8 GB go for it. Otherwise 2x 1060 are rlly good for the price.
Moved from “General Forums > Blender and CG Discussions” to “Support > Technical Support”
By 2x, do you mean SLI ? 1060 doesn’t work in SLI mode.
The price difference between 1060 and 1070 is massive (to me at least), so unless 1060 will always make me wish I got 1070 instead, I would go for 1060.
But I’d like to heart from folks who used 1060 and 1070 in production.
The 1060 has a bit better price/performance than the 1070, but 2GB extra headroom might serve you well, too. You’ll have to make that call depending on what you expect to do.
Blender Viewport currently doesn’t benefit much from a beefy GPU, so you should consider keeping the 680 in for display and use the 1060/1070 only for CUDA. That will keep your computer responsive while rendering, and it will also increase the amount RAM available for CUDA.
Multi-device CUDA rendering does not need SLI, in fact SLI should be disabled.
The newer APIs like D3D12/Vulkan also support multi-adapter rendering, which doesn’t need SLI either. The only game using this (that I’m aware of) so far is Ashes of the Singularity.
Interesting. So basically 2 GPUs in the system, one is used as single GPU normally, and when Blender uses GPU stuff, both GPUs are utilized and VRAM added together ?
I wonder if I can use 1060 as primary GPU (since I could use it for desktop VR) and leave 670 to add extra 2Gb of VRAM when rendering with Cycles.
Unfortunately that is not how it works. When rendering on multiple cards, the render data has to fit into every individual card’s VRAM. Therefore the card with the least amount of memory becomes the limiting factor.
So, when using one card with 8 GB and one card with 1 GB VRAM for rendering, the available VRAM will be 1 GB, not 9 GB. But of course you can use the 1060 for rendering and the 670 to run the monitor(s) to keep the user interface responsive.
Do you need it for your daily work/job? Go with the 1070 when not why would u care if the render takes 20 mins more?
if you’re on w10, know that 1GB is gonna get reserved for OS so you’ll have only 5GB to use
also nice to be aware of, when having two cards installed, even if not used as primary (display) card 1GB gets reserved - GTXs can’t be set as headless (computing only)
got 1060, but now i would rather go with 1070
Generally, for good performance, randomly accessed data needs to be resident in device memory. All the data that needs to be accessed as fast as possible must therefore be duplicated on all devices. Out of simplicity, many renderers therefore choose to not support “out-of-core” rendering at all, making the memory available a hard limitation.
For Cycles, this means all scene data is loaded into all devices at startup, whereas the tile data is per-device. I believe that Cycles now also supports spilling texture data into host (CPU) memory, which incurs a performance hit. Spilling data over to other GPUs wouldn’t really make sense on a PCIe interconnect.
Can you provide a source/reference for this?
In my recollection, in order to have responsive system by dedicating one GPU to CUDA, it is sufficient to not connect any monitors to it.
I’m not entirely sure this is accurate. I’ve burned through the 12gb of my titan on numerous occasions and it doesn’t appear to be reserving much of anything here, unless of course the system is hiding it from gpu-z.
Attachments
“Oh, what have i done, again.” :rolleyes:
-
10% of VRAM is reserved by win10.
VRAM usage in Windows 10 (RedShift Forum)
Low available vram on headless GPU (RedShift Forum)
Since am mainly used to 8+GB, i round 10% to ~1GB.
Got w10 on another machine and yesterday it updated itself to 1607, am so stressed about that OS… luckily the machine is only intended & used for experimentation and tests. With w7 all is good.
Sorry for not posting the source before
my bad 
Ah! Makes sense.