This paper describes aspects of the arithmetic algorithms, architecture, and VLSI engineering of the 64-bit floating-point unit of the T9000 Transputer. The unit is fully conformant to the IEEE-754 floating-point arithmetic standard, and has been implemented in a 1 $mu@m, triple-metal CMOS technology. The 160,000 transistor design performs addition in 40 ns, double precision multiplication in 60 ns, and double-precision division or square root in 300 ns. It will sustain 17 MFlops on the Linpac benchmark, yet occupies less than 15 mm$+2$/ of silicon - about 8.5% of the die area of T9000.
展开▼