Originally Posted by prjindigo
Vega is HBM memory bus *(HBM and HBM2 use same bus) and cannot be "cut down" because the data bus uses more than 80% of the pin interconnects on the chip. You can't "cut down" the HBM system to a 3092 or 2048... they'd have to be built that way to start with. There's a large amount of public ignorance about HBM and memory systems in general in graphics cards. The HBM system is a rigid solution with a buffer system between the ram and the GPU's processing.
I don't think I understand this. How is it that you can't make a cut down (salvaged) Fiji, for instance?
As far as I know, the HBM is not a factor here (as you cut CUs) but even if it is Fiji has 4 HBM bus segments (each 1024 wide) on the gpu die, so if you remove two HBM stack you will have enabled 2x1024 wide bus (and have two idling segments). Additionally, in HBM stack each channel is completely independent of one another, so each DRAM module can be disabled on the go (if needed) and it is supposed to be very flexible in case something happens.
As such I don't see a reason, why you can't make a cut down Fiji (Vega etc.)with enabled either 4096bit wide bus or 2048 wide bus etc.
Or do I need to go back to reading?Edited by C2H5OH - 6/15/16 at 5:42am