Signal integrity will probably be fine, you can always go with optical signalling for the long routes. What would be more of an issue is absurd complexity, latency from one end to the other, that kind of stuff. At some point, just breaking it down into a lot of semi-autonomous nodes in a cluster makes more sense. We kind of already started this with multi-core CPUs (and GPUs are essentially a lot of pretty dumb cores). The currently biggest CPUs all have a lot of cores, for a reason.