Originally Posted by Mahigan
Ugh, I think you're probably the only person who doesn't get what I'm saying.
Let me break this down for you:
1. Drivers control what goes into the framebuffer and what doesn't.
2. Driver optimisations can be used to store bandwidth sensitive data into the frame buffer and non-sensitive data into system memory.
3. HBM can thus be compared to a sort of embeded cache in the way AMDs drivers treat it.
4. The result is thus a 0 performance penalty from using system memory for non bandwidth sensitive data.
5. You get the full performance benefit of using HBM even if you're buffering more data than its capacity would allow.
Do you understand?
What the...? In my very first post to you, I directly pointed out and mentioned AMD's driver optimizations myself. You're not even comprehending what I'm posting, evidently. For the umpteenth time, what I'm questioning you on is this:
HBM is different than GDDR5. Its bandwidth is used like an embeded cache.
What about HBM or its bandwidth makes it different from GDDR5 in a way that makes it a better "embedded cache" or in any way different from GDDR5 in regards to capacity limitations? I've said it a million times already, but I'll answer it for you again. All VRAM is used for the purpose of storing local data that needs to be accessed quickly and all dGPUs, regardless of memory technology, can access system RAM at equal speeds for less immediately-necessary data.
HBM/GDDR5's bandwidth figures refer only to GPU <--> VRAM communication and don't do a thing to assist with VRAM capacity limitations (no particular memory capacity optimizations benefit from HBM's bandwidth) since all the data that must be stored locally will have to be regardless due to PCI-e limitations. A diagram with crudely-written bandwidth figures (512 GB/s for Fiji and HBM, 16 GB/s for PCI-e 3.0):
Data taken from system RAM into VRAM across PCI-e 3.0 is going to be that slow no matter what and therefore the same amount of data for any given scene needs to be stored in VRAM no matter whether the card is using HBM or GDDR5. Additionally, any optimizations involving prioritization and segmentation of data across VRAM and system RAM will be equally as effective regardless of GDDR5 or HBM for the same reason. VRAM bandwidth figures don't mean what some people, including yourself, think they mean. It has no bearing on capacity limitations because those limitations exist due to external factors regardless (like PCI-e).
HBM is not different from GDDR5 in any respect involving effective capacity or optimizations via the separation of data across the two primary memory areas.