•Thinking Machines Inc.’s MIMD machine [Caution: CM-1 and
CM-2 are SIMD]
•Goal: Create an architecture that could scale arbitrarily
•Nodes
are standard proc/fpu/mem/NIC
•Scaling
came in “powers of 2” using fat tree
•Special
hardware performed ‘reductions’
•“Programmed
I/O” meant PE was split between comp and comm duties