Motherboard: X470 ASUS ROG Crosshair VII Hero WIFI
CPU: Ryzen 2700x
GPU1: XFX RX Vega 64 Liquid.
GPU2: PowerColor RX Vega 56 Red Dragon.
RAM 4x16GB Corsair Vengeance RGB Pro.
HDD: Seagate 2TB
OS: Windows 10 64bit Pro version 19.03.
Has HBCC feature on Vega 64 and 56 GPUs been removed now that AMD are selling 16GB Radeon VII?
I have an RX Vega 64 Liquid Primary, RX Vega 56 Secondary.
Crossfire is off.
Both HBCC buttons are available for each GPU.
However when I reserve a 16GB HBCC memory segment, on each or either GPU it does not look like the memory is actually getting reserved in Windows 10 64bit Pro.
I still see almost full 64GB available.
Am I missing something?
It might be an AMD driver issue.
I cannot investigate further right now because I am running Blender 2.80 jobs at present and they need to be finished before the end of the week.
The HBCC switch must be doing something because when I do switch it on for either GPU, Blender 2.80 will not render anything at all using the Vega GPUs... I am assuming that is a bug in either Blender 2.80 or AMD Driver or both.
Is there any support for Blender 2.80 Cycles on AMD GPUs on this forum?
It looks like AMD are really starting to lag behind in Blender GPU rendering again, after a catch up with Nvidia in 2.79b - see: https://techgage.com/article/blender-2-80-viewport-rendering-performance/
Also I am seeing significant difference in CPU versus GPU rendering output results on Ryzen 2700X versus Vega 64/56.
One reason for having 8core 16 thread CPU is to use at least 12 threads for rendering, whist MultiGPU rendering on AMD Vega cards is running. However the quality of results are so different that I cannot use the CPU based results.
If anyone with an RX Vega 64 or 56 running Adrenalin 19.9.1 could look at how HBCC is behaving on Windows 10 64bit Pro I would be very grateful for some help.
Thanks.
Here you can see me running Blender 2.80 Multi GPU on the following AMD Cards:
Primary : XFX RX Vega 64 Liquid.
Secondary: PowerColor RX Vega 56 Red Dragon.
Tertiary : XFX R9 Fury Triple Dissipation.
In the screenshot you can see 3 tiles processing on the left.
On the right hand side I am running 12 of 16 available threads of the Ryzen Processor.
The "Remaining" run times are just estimates.
I will update runtime information next.
The Radeon Settings (19.9.1) shows that I had to turn off HBCC memory Segment on both of the RX Vega 64 Liquid and the PowerColor Red Dragon Vega 56 to get Blender 2.80 to run at all.
Question 1 is this.
Why does Blender 2.80 not work with HBCC option on RX Vega 64 and 56 GPUs?
Question 2 will follow later - but here it is. Why does the CPU render look so different to teh GPU render in AMD OpenCL MultiGPU Rendering versus the CPU Rendering results?
Exact same settings are used for both Blender runs initially.
I am looking at Blender GPU versus CPU Redering differences - I will post linj to that information next.
Question 3. What does a MultiGPU / GPU render look like on Nvidia CUDA versus the CPU rendering result.
Thanks.
I have asked about HBCC support on Blender Stack Exchange here:
https://blender.stackexchange.com/questions/127258/vega-hbcc-support/152591#152591
Hey! I have a vega FE so I might be able to answer some questions. Disclaimer, this is mostly my memory and personal experience with one that REALLY NEEDS NEW THERMAL COMPOUND
Q: Has HBCC feature on Vega 64 and 56 GPUs been removed now that AMD are selling 16GB Radeon VII?
I have an RX Vega 64 Liquid Primary, RX Vega 56 Secondary.
A: No, it's still there, even now.
Q: Both HBCC buttons are available for each GPU.
However when I reserve a 16GB HBCC memory segment, on each or either GPU it does not look like the memory is actually getting reserved in Windows 10 64bit Pro.
I still see almost full 64GB available.
Am I missing something?
A: You're not missing anything. HBCC is a way to more effiently allocate GPU memory by using some system RAM to store graphics data. It does not steal any system memory though as its allocation is dynamic (within the limitation you set and your system)
It's not a driver issue.
Blender has always had problems on AMD for no apparent reason, even when using the HIP/open-cl backend. (multi-device is broken too, and HIP-RT uses way too much memory)
so TLDR: your HBCC is working as intended. HBCC doesn't hard allocate a portion of system ram to the GPU, instead it dynamically moves assets between system RAM and VRAM to increase efficiency (windows will show the HBCC memory amount as your VRAM, but other apps will show the actual hardware vram). HBCC's memory allocation is dynamic, and is designed to make VRAM allocation and usage more efficent by also using system memory for things.
I have noticed slight performance uplifts when using HBCC