blender-2.78-94a3bc6-win64, Win10
Latest drivers
This is a default scene with sphere:
Viewport rendering works fine altough it takes long to open render kernels:
Viewport rendering on CPU:
On GPU memory usage is about 585 MB and on CPU is only 10MB.
After I open viewport rendering again, this error happens:
The AMD HD 7870 is only a GCN 1.0 card for 2.79 we bumped our requirement to GCN 2.0. For the final release, these cards will not be selectable from the menu.
Please, could you leave the compatibility requirement for 7870?
Maybe someday AMD will change the drivers so that issue will be fixed. I have been writing the feedback to them often.
No this is on the hardware level and not software so the GCN (Graphics Core Next) version will not change.
You can view what GCN a card is on https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units
Why are you leaving GCN 1.0 behind? I think this is some kind of memory leak issue.
First render on GPU has memory usage about 585 MB and on CPU is only 10MB. Shouldn't this be the same ?
I believe second render gives an error because memory allocation is stacked.
We are leaving GCN 1.0 behind because of various errors like this that we cannot fix. @Mai Lavelle (maiself) may wish to comment further is she would like.
@Aaron Carlisle (Blendify) , @Mai Lavelle (maiself)
On this thread https://community.amd.com/message/2772700 there is a confirmation that ticket has been raised against this issue with 7870. It is being investigated.
Please, don't remove 7870 from 2.79 if anyhow possible.
We will be dropping support for GCN1 because we cant fix these issues for a variety of reasons:
Its very unfortunate that we have to drop support for those cards, but there's really nothing we can do about it. If AMD manages to fix all the issues we could enable them again, but there's no telling if or when that would happen, so for 2.79 they will be disabled.
As for the memory usage, it is expected. The way the split kernel works is it tries to render 100s of 1000s of rays concurrently, each ray kept in memory to pass from kernel to kernel, each of which processes a different part of rendering the rays. Each ray state is about 2kb-5kb in size (depending on the scene), so you have something like 2kb*(100000..1000000rays) = 200mb .. 2gb of storage. Cycles will detect the largest block of available memory on the device and will create just the right number of rays to fit into that memory, so even tho it looks like a lot it will never affect the ability to render a scene.