Today we had an overview of the history of architecture, instruction set architecture, microarchitecture, the diversification of architectures, and how architects sometimes fall into the trap of claiming that improvements in fabrication technology and obvious changes that result from scaling, as well as improvements in the software stack, are due to the cleverness of new architectural features. We then saw how the annual conference publication paradigm can lead to the cutting of corners in good research methology, especially when simulation of new systems is both very time consuming and also requires changes in the software development tools. Actual fabrication of new hardware is very costly but, as we will see shortly, often reveals issues that simulation didn’t capture, and that affect performance. We then reviewed the paper by Diwan, and discussed some of the implications for his finding that many reported results actuall fall into the noise of common environmental variations, and that setup randomization is needed, even though it increases the effort. 

Slides are here