Hewlett-Packard Co introduces its first 64-bit microprocessor, the Precision Architecture RISC 8000 this week, reports today’s edition of our sister publication Unigram.X. It rates the part at 360 SPECint92 and 550 SPECfp92 when clocked at under 200MHz. When the bi-endian PA-8000 ships in systems – which are promised by this time next year – Hewlett-Packard believes that unprocessors will do 700 transactions per second and four-processor models 3,000 tps, to outgun the PowerPC 620, UltraSparc, Alpha 21164 and R10000 designs that will be around by then. As expected of the first all new Precision Architecture design since 1991 – hence PA 2.0 – Hewlett-Packard is cramming more extensions into the PA-8000 than any other RISC vendor has done so far. The integrated PA-8000 processes four instructions per cycle, performs out-of-order execution, has two floating point units, and load and store pipes. As with PA 1.0, there’s no on-chip cache, but up to 4Mb data and 4Mb instruction cache is accessed by a 128-bit wide interconnect. The 0.5 micron CMOS 8000 is taped out and will be followed in 12 to 18 months’ time by a design shrink to 0.35 micron, and tweak, to create the 8200. The 8000 and 8200 are expected to turn up in systems with up to 16 CPUs – Hewlett-Packard currently offers a 12-way top end. Clustering takes over thereafter, it says. The 8200 will take Hewlett-Packard up to the PA-9000, originally the 1998 vehicle for the first of the technology developed jointly under its alliance with Intel Corp. However, while Intel put its picture into better focus last week, saying the 1997 P7 will bear first fruit of the agreement, Hewlett-Packard’s vision appears a little fuzzy. It says it may introduce some of the jointly-developed features in p re-PA-9000 technology, although it hasn’t yet decided whether the 9000 will include fully-integrated iAPX-86 functionality, or only some components of what the two have done by then. Meanwhile, it seems the techniques Hewlett-Packard and Intel are using should be better characterised as the bastard son of VLIW than Very Long Instruction Word. We shouldn’t call it VLIW, Hewlett-Packard told us, but admitted it doesn’t know what to call it yet. Whichever way it turns out, Hewlett-Packard’s performance targets call for the PA-9000 to deliver at least 1,000 SPECint92, or 10 times the performance of the company’s current the PA-7100 RISC-based family of machines.
