Function, latency, bandwidth, power: towards a better computer

Computation in practice consists of executing functions, usually with the intention of achieving all three of correctness, high throughput, and low latency. Unfortunately, the most popular ways to specify functions - e.g., C++, Java, etc. - have not advanced significantly in 20 years. Mainstream software is hard to reason about and lacks any means of even expressing throughput or latency requirements, much less satisfying them. On the hardware front, the throughput of uniprocessors has barely increased in 10 years, yet developing efficient algorithms for multi-core processors remains ad hoc and frequently fails. Latency is no longer improving either -- because multiple cores now compete for limited, off-chip memory bandwidth, and each core needs access to lots of off-chip memory. Finally, the ever-increasing need for communications throughput for datacenters, the Internet, and mobile devices is at odds with the importance of decreasing power consumption. What is a supercomputing community to do? (Hint: I believe that it is neither multi-core microprocessors nor GPUs.)