Originally Posted by Usario
This is not a fully enabled part, it has 5120 SPs and a 3072-bit memory bus. GV100 has 5376 SPs and 4096-bit.
You can count on yields being crap with an 815mm^2 die, interposer and HBM2.
No, you're wrong in several ways.
The first of which is that this is a damaged bin
part not an under-enabled part. nV can charge 2x as much for fully functional parts. These cards have a bad HBM2 stack on them or they'd be 16gb.
The second is that each HBM2 stack is only 128 bits wide and cannot be addressed with larger than 64bit data, it has two 64bit channels into it not 1024.
The parallelism of the ram in
the HBM2 stack is immaterial.
These cards are 384bit, are missing an HBM2 stack, may have damaged nVlink parts and are useless in arrays because of it.
I've described elsewhere how HBM and HBM2 stacks work but basically the "ram controller" for the stack is in the bottom and has access channels to the PCIe and the GPU, the GPU communicates with 2x64bit bi-directional channels per stack at very high speed.
Claiming that the HBM is 3072 or 4096 wide is like claiming that 4-channel boards with 8 slots populated with 8gb sticks of ram are 1024bit wide memory buses. There may be 1024 bits of memory port address in there but it's still only 4x64 bit channels to the CPU for 256.
Don't propagate marketing twaddle like "4096" and "x86"... they're lies.