I am curious if Vega will require more system memory to run properly compared to regular cards due to its paging techniques? Are there any tech details on this yet?
Obviously yes. I mean, if you have a 4GB card and have a game with 6GB worth of textures, the data has to live somewhere, doesn't it?
We still don't know anything about how the HBCC works, so one can only guess. Here, let me go wild:
Maybe it would be possible to have a slider in Wattman for configuring how much "total VRAM" you want. If you set it to 8GB (and have a 4GB card) you'd need an extra 4GB from system RAM, but that'd be up to you. Say you want to use only 2GB of system RAM maximum, then you could just set the slider to 6GB and you'd achieve that.
Maybe games themselves would be able to configure this by talking to the driver (some AMD-specific API) so that they negotiate the best amount of system RAM to use, based on their actual texture pack size, the actual VRAM on the card, and the actual available system RAM.
Either way, actual usage will be 'whatever the game really loads'. I don't know how technically-minded you are, but there is such a thing as "virtual address space" for memory (which Vega's HBCC is heavily using) and that allows you to translate a memory address to something physical (RAM on the card OR in the system), but it can also be 'orphan' and translate to 'nowhere' (we call such an address unmapped). So if you don't actually load anything into that virtual address, it will not actually occupy actual RAM. Example again: I have the 4GB card, configured it as 8GB in Wattman, but am running a game that actually uses 4.5GB: the game would start loading into the card's virtual addresses and anything from 4GB to 4.5GB would end up occupying half a gig of system RAM, but that's all that would be used (in that sense, the 8GB I configured in Wattman is a limit, not a buffer to be reserved).
Maybe Vega supports compression to move the data in packed format and better exploit bandwidth (like they already do on the cards themselves). Sure for actual VRAM this does not allow you to save space (it's only done to increase throughput), but in the HBCC case it actually can also save space (due to virtual addresses as explained above). So if a game is using 6GB of which 2GB need to be in system memory, compressing them may end up making those use fit in just 1GB.
Again, we know nothing and the range of options is limitless!