Again, where are you getting 50-100MB from? you keeping throwing out these numbers, but have so far provided zero evidence to back them up.
And regarding the other sources of latency, what about them? RAM and interface latency is generally in the nanosecond range and will thus have zero impact on this issue.
I can't believe I'm talking to a person who doesn't understand the difference between frame latency and frame interval:
But maybe you think Nvidia is also crazy when they claim a 50ms total frame latency at 60FPS:
As you can see there is a 32ms gap from when the game engine thread finishes its work on the relevant frame until the GPU finishes it's work. If we say the GPU takes 10ms or so before it gets to the texturing step (since that is generally one of the later things in the graphics pipeline), then that gives you 25 ms to transmit your necessary data. Transmitting 5-20MB over PCIe would only takes about 0.5-1.5 ms, so over an order of magnitude less time than what is available. Obviously you also have system RAM, VRAM and PCIe packet latency on top of this, but as mentioned these should all be in the nanosecond range, and thus hardly relevant.
If you actually bothered reading my posts you would see that this came from Sebbbi from B3D, I even provided a link for you.
If you have a game where you previously need 2-3GB*, then yes you could expand the example to a 1GB GPU, or a game which only used 512-768MB of VRAM could be handled by 256MB VRAM.
Of course games today often use 6-8GB of VRAM with the current memory management system, and as such you would need about 3GB with the new system, 4GB to be safe and for a bit of future proofing.
Sigh, No one is saying that you can get everything the frame needs over PCIe during rendering, people are saying that you can get all of the new data needed over PCIe during rendering. Yes GPUs can have half a terabyte of BW, or ~5GB per frame at 100FPS, but you don't need 5GB of new data every single frame, your only need the aforementioned 5-20MB (since all of the other necessary data can be reused from the previous frame). Do you seriously think the GPU flushes out every single piece of data once it finishes a frame, thus requiring you to fetch all of the relevant data for the next frame anew?
The reason why GPU VRAM bandwidth is so high, is due to the things it has to access other than just asset data, things like render targets, buffers, and other temporary data.
Obviously you need to load in a bunch of stuff for the very first frame since this frame requires a lot more than just 5-20MB of new data (since all of the necessary data is new, when it's the first frame), and PCIe is of course a bottleneck here, which is of course also why games have loading times when you first start them up or move to a new level.
I like how you accuse people of "attacking the messenger", when in the previous paragraph you just did that yourself ("Who told you that? What games has that expert shipped?")
And how in the world am I wrong about hUMA, when I haven't even mentioned hUMA?
And what calculations are you talking about exactly, and why do they have errors of 10MB/frame? You keep on throwing out all of these claims and yet provide zero evidence for any of them.
And as I have explained to you repeatedly now, the order of magnitude difference in bandwidth/latency between PCIe/system memory and VRAM isn't relevant for this discussion (about HBCC), since they are used/accessed in completely different manners.