Abstract: A node controller (12) includes a processor interface unit (24), a crossbar unit (26), and a memory directory interface unit (22). Request and reply messages pass from the processor interface unit (24) to the crossbar unit (26) through a processor interface output queue (52). The processor interface unit (24) writes a request message into the processor interface output queue (52) using a processor interface clock to latch a write address from a write address latch (62) in a synchronizer (60). The write address is encoded by a Gray code counter (64) and latched by a first sync latch (66) and a second sync latch (18) using a core clock of the crossbar unit (30). The output of the second sync latch (68) provides one of the inputs to a read address latch (70) using the core clock of the crossbar unit (30).

Abstract: The present invention optimizes the number and ratio of cycles required among the divide/square root unit, multiplier unit and ALU. An intermediate latch with its own clock is provided at the output of the multiplier half-array in the intermediate stage to feed back data for a second pass for double-precision numbers. The multiplier can then be adjusted for either two-cycle latency mode (for optimizing double-precision multiplies) or three-cycle latency mode (for optimizing single-precision multiplies). A separate divide clock is used for the divide/square root unit, and is synchronized with the multiplier cycle clock on input and output. This allows the divide time to be optimized so that it requires fewer clock cycles when a longer multiplier clock cycle time is used.