Originally Posted by Mahigan
It's possible to do on GDDR5 but GDDR5 isn't as fast as HBM.
Which is completely irrelevant to the ability of a GPU's VRAM pool to store or swap in/out data from system RAM. HBM-based bandwidth figures refer only to the speed of the GPU in reading from and writing to VRAM. VRAM exists as local storage because system RAM is too far away/slow (because of PCI-e limitations, which are completely unaffected by VRAM bandwidth) to capably store all the data any given GPU needs at any given time.
Whether it's Fiji's 512 GB/s, GM200's 336 GB/s, GM206's 112 GB/s, etc. for example is completely irrelevant to accessing system RAM and therefore capacity remains a comparable limitation (meaning no additional ways to get around it) regardless of VRAM bandwidth. Therefore, your claim of HBM being different doesn't strike me as relevant.
And I never said it reduced the data held in system RAM. You created that strawman yourself.
I created that? I never even said that. I was disagreeing with your position that it made capacity in VRAM
non-comparable across GPUs with differing VRAM bandwidth however, which is the only possible thing you could have been implying with this statement, particularly the section in bold:
HBM is different than GDDR5. Its bandwidth is used like an embeded cache. Anything that requires quick access, sits in the HBM, other lower bandwidth sensitive storage needs go into the System RAM.
Implying that VRAM capacity (and/or associated limitations) between HBM and GDDR5 isn't entirely comparable because, "HBM is different than GDDR5" is the only relevant thing you could have meant by this.
Yes it is 4GB capacity but used as a quick access embeded cache like we see on some consoles. Any data requiring quicker access resides on the HBM, data which is not bandwidth bound resides on the system RAM. The end result is 0 performance hit.
Excluding whether or not you're examining enough cases to say that is the case that there is "0 performance hit", please explain what you mean in the context of your above quote about HBM being different? What about HBM being "different" from GDDR5 has any bearing on capacity limitation comparisons?
Prior to the Fiji launch AMD had this to say:
And they were right.
Right that certain optimizations could be made, of course. Right that Fiji's VRAM would never be a problem? Haven't seen enough actual testing on the matter to figure it out; VRAM in general isn't very strenuously-tested in benchmarks, but that's irrelevant to my contention of your point of HBM being different.Edited by Serandur - 3/1/16 at 6:44am