A big part of the performance of these machines came from high bandwidth memory access. This is high performance more in the sense of an 18-wheeler than in the

A big part of the performance of these machines came from high bandwidth memory access. This is high performance more in the sense of an 18-wheeler than in the sense of an F1 car. The point is more about predictable control flow and bulk manipulation of data than it is lots of tricky control flow, etc. Of course, this analogy isn't really complete in that it misses the fact that Cray's machines also pushed the limits on cycle times. As his career evolved, his focus shifted from circuit design to reduction of propagation delay with more tightly packaged machines, then to cooling those tightly packaged machines, and then ultimately to GaAs as a way to get switching times down. So not just an 18-wheeler, but one that happens to run at 200 mph, with all sorts of custom high-end engineering to make that happen with technologies of the era. The fact that these days, we can emulate this with programmable hardware is an amazing testament to exactly how much hardware has improved over the years. Something else I should mention is that Cray also had an element of design conservatism to his work. The Cray 1 is famous for being a vector supercomputer, but it was not the first - his former employer, CDC, beat him to the punch with the CDC Star. Unfortunately for CDC, they got the balance of the design wrong. The Star wound up with vector capabilities that took too long to set up to be performant and scalar performance that was too slow to make up the gap. Cray's design for the Cray 1 got around both of these limitations, which is part of why it was as successful as it was.