CRAY

In 1976, the CRAY 1 was announced, with an 80 MHz clock cycle. This single processor machine was based on ECL technology, and was able to read from and store to memory one word per cycle (a 1 word memory bottleneck). In '82, the X-MP was announced, a 2 (later 4) processor machine with shared memory and 120 MHz clock. An important breakthrough with the X-MP was that 2 memory loads and 1 memory store were available each cycle. This kind of memory access is still the widest bandwidth that has been built. In 1988 the 8 processor Y-MP was delivered, followed soon by the C90, a 16 processor, 240MHz machine. It was at this time that Cray's financial troubles put the company's future in some doubt. There followed in a short time the Cray J-90, based on CMOS technology, and the T90, an ECL machine requiring liquid immersion cooling.

For all these machine architectures, Cray developed 64 bit floating point precision and 64 bit word addressing, efficient compilers, vectorization capabilities and compiler preprocessing that take advantage of vectorization and, for multiprocessor machines, parallelization via compiler directives.

In 1996, (part of) Cray was taken over by SGI. Leveraging the best of these two companies, the Origin 2000, a ccNUMA machine, was introduced in '97. ccNUMA, cache coherent, non-uniform memory access, is a version of shared memory. Every processor "owns" and has very fast access to its own memory. Every processor also has a copy of the global memory space, and can access memory "on" another processor. Depending on where this datum resides, that access may be over a single switch to a nearby processor's memory, or it may be over several jumps to the memory of a topologically remote processor. Cache coherence means that memory access is temporarily blocked if a procesor is operating on its own memory (thus in local cache), until the operation is completed. In this way, data is accessed in a consistent manner - the picture of memory is coherent.