Cydra 5 supercomputer: design philosophies, decisions, tradeoffs (1989) [pdf]

The actual title is The Cydra 5 departmental supercomputer: design philosophies, decisions, and trade-offs. I had to shrink it to fit.

If you're familiar with Itanium, then the design will look familiar. HP hired some of the Cydra designers.

"One of the more ad hoc decisions we made was choosing the number of registers per register file. The problem we faced was a lack of directly relevant statistics on the effect of this parameter on performance in the context of a directed-dataflow style of execution. So we plucked the decision out of thin air. Our collective intuition told us that 32 registers per register file was too few, and 128 registers looked difficult from an implementation viewpoint. Thus we settled on 64 registers per register file. Comforted by the lack of alternatives, we moved on."

The late 80s/early 90s was such an interesting time for supercomputer architectures. I don't remember ever hearing about Cydra. But there were lots of niche ones -- Kendall Square Research had the ALLCACHE architecture, there was the Connection Machine from Thinking Machines, and the Tera MTA with lots of threads to hide memory latency. There are lots of others (anybody have an interesting or favorite one I'm forgetting?)

Interesting stuff.