Vega Professional cards have been out for a few weeks now. We know the High Bandwith Cache Controller can offload VRAM objects into slower memory so that the GPU doesn't run out of available memory. Would be nice to see this integrated into Blender rendering. The GPU using AMD's HBCC would probably be slower than straight VRAM, but still way faster than a CPU and you could use system RAM which is way cheaper.
j
jason
Do you have any numbers to support the "still faster than CPU" claim? I can see the value of offloading to system memory in order to prevent failures and crashes, but I'm dubious that the result is necessarily guaranteed to still be faster than straight CPU processing. I/O bottlenecks can be pretty time-consuming.
j
johnsonjp34Author
I don't have numbers, but Bethesda and other large game studios claim that they will be making vastly larger worlds in their games because of the HBCC ability to sort memory objects as needed.
They haven't said what the latency will be though using the Vega memory controller with system memory though.
j
jason
Just be aware that the way memory gets used for GPU compute tasks (like rendering) can be pretty different than how it gets used in real-time environments like games. As a result, I'm not sure how likely it is that supporting this feature would provide any measurable performance improvement (aside from, as I mentioned before, not crashing/failing on large scenes).
4 Comments
Vega Professional cards have been out for a few weeks now. We know the High Bandwith Cache Controller can offload VRAM objects into slower memory so that the GPU doesn't run out of available memory. Would be nice to see this integrated into Blender rendering. The GPU using AMD's HBCC would probably be slower than straight VRAM, but still way faster than a CPU and you could use system RAM which is way cheaper.
Do you have any numbers to support the "still faster than CPU" claim? I can see the value of offloading to system memory in order to prevent failures and crashes, but I'm dubious that the result is necessarily guaranteed to still be faster than straight CPU processing. I/O bottlenecks can be pretty time-consuming.
I don't have numbers, but Bethesda and other large game studios claim that they will be making vastly larger worlds in their games because of the HBCC ability to sort memory objects as needed.
You are correct about IO bottlenecks. That's why AMD will be releasing the 2TB card: pro.radeon.com/en-us/product/pro-series/radeon-pro-ssg/
They haven't said what the latency will be though using the Vega memory controller with system memory though.
Just be aware that the way memory gets used for GPU compute tasks (like rendering) can be pretty different than how it gets used in real-time environments like games. As a result, I'm not sure how likely it is that supporting this feature would provide any measurable performance improvement (aside from, as I mentioned before, not crashing/failing on large scenes).