With Dennard Scaling gone, and Moore's Law almost gone, if computer performance is to continue to improve moving forward, more will be demanded of computer architects...and that means doing research that moves the needle. With 60.5 years in this business, I have some thoughts as to how to move the needle. First, work on challenging problems. Mickey mouse does not move the needle. Understand the transformation hierarchy, and take to heart the MIT white paper reminding us that there is plenty of room at the top, ...and my comment that there is still plenty of room at the bottom. It is no longer okay to be just a hardware person or just a software person. Embrace heterogeneity. It can simultaneously provide improvements in both performance and energy consumption. Recognize that there are often more than one bottleneck and you will not see a big win until the last bottleneck is removed. Recognize that energy is a first class citizen that often negates the value of a performance win if the energy cost is too high. Hardware security needs to be addressed. The microprocessor of 2035 will be dominated by accelerators which will need to talk to each other. Not all will be active at the same time. That bubbles up two things: serious communication and something to maintain order, rather than letting chaos prevail. To maintain order we will need a von Neumann machine; contrary to popular whimsy, the von Neumann machine is not dead, and won't be as long as we need accelerators to coordinate with each other, and it makes no sense to power up all of them at the same time. Although economics drives business models, and maintaining compatibility is important to economics, don't be afraid to throw compatibility out the window if it helps solve a problem that is so important that economics should not drive your research. This talk will look at most of the above.
Patt Patt (Fri,) studied this question.