With a mesh it's probably ok so long as the mesh is always clocked higher. Links would generally be idle with no contention. Still a weird design as I thought they said it was connecting GPU clusters. This is more like a GPU with some other components attached versus an MCM style chip like Ryzen.
AMD did not say much beyond indicating Vega had the fabric implemented as a mesh, and that it matched the full bandwidth of the memory controller.
Indicating that its bandwidth is on the order of the HBM interface gives a general ceiling and floor to its bandwidth.
I had doubts it would integrate between the CUs and L2 because it would have been a massive drop in bandwidth.
At 16 slices and say 1.7 GHz, the mesh would have strangled the CU array with bandwidth 3.4x too low.
Looking at the strip of hardware that appears to match AMD's diagrams for the IF section, it runs across one side of the GPU between the ROPs and HBCC. There appears to be one rectangle per channel, which may make that one of the broader sides of the mesh, with the L2 segments being on the other. The remainder could hang off the narrow end.
Creating enough blocks to handle 64 CUs would have plugged multiple times that area right in the middle of the GPU.
Have we some knowledge about the efficiency of Infinity Fabric links ? Is this a good technology, or is it hell to manufacture chips with this tech in it, how much power does it need for very high bandwidth need like Vega, etc ? Can Infinity Fabric tech be what's wrong with Vega ?
Infinity Fabric uses a superset of Hypertransport, which has 8-12 bytes of command overhead per 64 byte transfer. The fabric likely has more complexity in its individual switch points, given its variable topology composed of point to point links. The independent clocking likely injections some latency (shouldn't matter that much in this instance).
I would assume this is all measurably higher-overhead than the straight and generally dumb direct link from past GPUs, although in a forward-looking sense it can provide more options than said dumb link.
The overhead aside, it seems like the fabric is robust enough, and its protocol and implementation should be well-understood given its mature roots. At least in theory, Vega 10 shouldn't be really stretching many of its capabilities.
