There are some computers that that are legendary for their revolutionary architecture and design. One indisputable member of this category is the legendary Connection Machine CM2 from Thinking Machines. I own one of the original 1987 vintage CM2 machines. I have future plans for this but may consider selling it given the right circumstances. Please contact me if you're interested.
This unusual machine has many small, limited processors that work together in lockstep. A central unit broadcasts a command to each, and they all execute together. The best-known data-parallel machine is the 1987-vintage, 64,000-processor CM2 from Thinking Machines.
- From Byte Magazine article from February 1994 / State Of The Art / Types of Parallel Machines
Connection machine CM2
Architecture: The machines consist of an array of simple proprietary bit-serial processors directly connected to local memory. Groups of 32 processors may optionally share a floating point accelerator unit. The memory uses commercially available dynamic RAM. The machine is driven by a Sun, Vax or Symbolics 3600-series machine.
Node: The basic building block is a two processor chip pair (32 Processing Elements) and an optional floating point accelerator.
Topology: The nodes are connected as a hypercube.
Operating System: Whatever runs on the front-end.
Languages: CM Fortran, C* and *Lisp.
Programming Environment: TMC have a powerful window based development environment known as PRISM which offers debugging, datatype visualization and monitoring.
Performance: The manufacturers quote a peak performance of 28 GFLOPS for the largest CM-2 (65536 PEs) with floating point accelerators. The sustainable rate is about 5 -- 10~\% of this figure, the highest achieved being 5.6 GFLOPS.
Data Transfer: Using the NEWS grid for near-neighbour communications of dimension less than 16, the manufacturers give a bandwidth figure of 2.5 Gbytes/second for a 65536 PE CM-2. The global router achieves 250 Mbytes/second. An octant (8192 PEs) has two I/O connections which can be linked to a disc-farm (the DataVault), or frame buffer. Speeds of 50 Mbytes/second are possible, with one channel active at a time. The DataVault can sustain 25 MByte/sec and striping can improve this speed.
Scalability: The CM2a has 4096 or 8192 processing elements (PEs). The CM-2 has 16384, 32768 or 65536 PEs.