Exploiting Distributed Shared Memory in GPUs for Next-Gen AI workloads