Can custom RISC chips remain viable?
> I thought MIPS processors were supposed to be more streamlined for
> specialized processing which gave them a great advantage over
> Intel. (I have a knowledge gap on microprocessors) Is that why
> the Cray microprocessors are integrated into the high end SGI
> boxes??
Intel processors are CISC (Complex Instruction Set Computer) processors. UNIX workstations use RISC (Reduced Instruction Set Computer) processors. As its name implies, RISC processors used to have a smaller instruction set than their CISC cousins. RISC processors used the extra silicon real estate (that CISC uses for the extra instructions) to add internal caches, lots of registers, and very powerful floating point units. However, the PowerPC, a RISC processor, has nearly as many instructions on chip as a 486! Many of these instructions are specialized for signal processing applications, not general computing. In today's market, that's how RISC processors attempt to distinguish themselves from each other. They add specialized instruction sets or boost floating point performance. If the small instruction set no longer distinguishes RISC from CISC, then what does? RISC processors use a fixed length for instructions which makes them easier to manage and optimize performance for. CISC processors are forced to use variable length instructions because the more complex instructions require more data arguments than the simpler ones. On-chip optimizations for CISC processors are inherintly more complex and require more silicon real-estate. RISC processors also tend to have better floating point performance and still have more registers than Intel chips. However, the Pentium and Pentium Pro processors have shown that Intel wishes to challenge RISC processors dominance of floating point performance. The next generation Intel processor promises to be a marriage between CISC and RISC technology. Intel has teamed with Hewlett Packard, a major RISC player, to design the chip, code named Merced. They are very tight lipped about the chip's design. The only thing certain is that it will be a 64 bit processor. All else said about this chip is rumor and speculation. In all likelyhood, it will be a chip with a RISC core that emulates the more complex x86 instructions by breaking them into simpler RISC instructions. It will probably also use some scheme to increase the number of registers. This will be CISC's last breath. The next Intel processors will probably be purely RISC. The transition should be relatively painless. Even today, software made for Pentium and Pentium Pro processors uses only a small portion of the x86 instruction set for performance reasons. Those complex x86 instructions that boosted performance in older generation chips only hinder performance in the new generation of Intel chips. Most compilers don't generate those instructions anymore; they only turn up in hand coded assembly routines.
The real nail in the coffin for custom RISC chips, however, is the prohibitive cost of producing the next generation chip. Upgrading an existing fabrication plant costs $500 million dollars; building a new one costs billions. The only reason that custom RISC chips remain competative today is because the chips are produced by third parties with modern fabrication facilities used to produce a variety of chips. Neither SGI nor SUN actually make their own processors. If I remember correctly, there are fourteen different manufactures of SPARC chips. The largest is Texas Instruments; they make the chips in the same plant that they use to produce high speed DSPs (Digital Signal Processors). However, even this is a stop-gap solution. Eventually, new fabs will get so expensive that recalibrating a plant to make a few thousand RISC chips won't make economic sense; the chips would have to be sold at an unjustifiably high price (compared to high volume Intel chips).
HP will apparently be the first UNIX vendor to use an Intel chip for workstations. Others UNIX vendors will have to follow HP or put their support behind a single alternative. Right now, the only viable alternatives are the PowerPC and MIPS. They are the only RISC chips produced in relatively large volumes. PowerPC has the edge because it is already produced in large volumes for computers and workstations, namely Apple computers and IBM workstations. MIPS volume mostly ends up in game consoles. Moreover, I speculate that SUN is secretly considering a move to the PowerPC. Motorola is porting (or has ported) Solaris to the PowerPC. It would be the move with the least risk and the most culturally correct for SUN. Does this damage SGI in any way right now? No. This transition is still a years away. It only hurts SGI if they stubornly hang-on to MIPS after it no longer becomes viable economically. By then, MIPS should have more than payed for itself. |
