Sun Microsystems Inc was due this week to describe its answer to the 64-bit MIPS Computer Systems Inc R4000 RISC chip with details of the SuperSparc, the Viking part it has been working on with Texas Instruments Inc. Sun describes the chip as a fully integrated single-chip superscalar Sparc RISC, fully compatible with its predecessors and outperforming the Sparcstation 1 by a factor of from five to 10 times running existing Sparc software using existing compilers – that would put performance at from 60 MIPS up, though 80 MIPS has been most often mentioned. Future compilers should wring additional performance out of it. It will be touted to be priced appropriately for desktop to superserver configurations. Sun is expected to use it in its upcoming Galaxy multiprocessors, and some Sparcsystem builders may be using the part in machines of their own. SuperSparc uses the Texas 0.8 micron BiCMOS process, enabling it to squeeze 3.1m transistors on the chip. It is said to integrate integer, floating-point, memory manager and multi-processing MBus on-chip with cache, executing up to three instructions during a single clock cycle to produce a serious level of parallelism over serial execution. Unlike the MIPS R4000, which claims to be the first true 64-bit part with 128-bit data paths based on a 32-bit architecture, it will not be full 64-bit, Sun saying there is no sense in paying for slower 64-bit arithmetic and pointers until they are truly needed; it does allow for double-precision 64-bit floating point. The design team claims that superscalar techniques are preferable to the MIPS pipelining approach because a superpipeline processor stalls frequently, as it cannot group and schedule instructions, and also that it is superior to the simple pipelining used by Hewlett-Packard Co in that it is more cost-effective, and longer lived because its peak performance starts higher.
