Someone should make a discrete gpu with an arm cpu core or 50 sharing the same memory and maybe ssd. Then you have a really nice hierarchy of parallelism, 2 fas

Someone should make a discrete gpu with an arm cpu core or 50 sharing the same memory and maybe ssd. Then you have a really nice hierarchy of parallelism, 2 fast arm cores, 10 medium, 50 slow, then gpu. Would be great for stuff like multigrid methods where you are balancing parallelism and latency: https://devblogs.nvidia.com/parallelforall/high-performance-geometric-multi-grid-gpu-acceleration/ https://devblogs.nvidia.com/parallelforall/high-performance-...