Getting CISC into RISC

Bob Ryan

The RISC-versus-CISC debate is taking some interesting turns that could have a profound effect on the future of the personal computer industry. For the past few years, conventional wisdom has stated that, all things being equal, RISC performs better.

What's muddying the waters is software. For better or for worse, as Andy Keane, manager of product marketing at Mips Technologies, puts it, the 80x86 architecture ``owns the binary.'' The one bright spot in the RISC firmament--Apple's success with its Power Macs--is because of Apple's powerful influence in the Mac software universe.

The challenge facing RISC vendors is to support enough of the 80x86 software base so that switching to RISC becomes a tolerable alternative to continuing to invest in 80x86 machines. The prevalent approach today is to offer some sort of software emulation that allows users to continue to use their DOS/Windows software on RISC hardware. But software emulation, for the most part, is too slow, while the installed base of RISC PCs is too small to tempt most software makers into converting their 80x86 software into native RISC code.

A more patient approach is to wait for the binary to come to you. Mips and DEC are banking on Windows NT to provide a ready source of applications that can be easily ported to RISC. The fact that an application written to the Windows NT Win32 API can be compiled relatively painlessly to 80x86, Mips, Alpha, and (soon) PowerPC platforms encourages the likes of Mips, DEC, and Motorola to stay the course while NT increases its market share. Their RISC PC strategies are predicated on the belief that the migration of the Windows 3.1 installed base to the far-more-capable NT platform is inevitable.

Chicago (aka Windows 4.0) applications written to the Win32 API should increase the number of p rograms that can take advantage of most of the features of an NT platform. But Chicago's popularity may discourage faster migration to NT, especially given that NT currently requires about three times as much RAM as today's 4-MB PCs. Neither Mips nor DEC sees a major impact for NT in the general desktop market before 1996.

Two years is a long time, and a number of companies are working hard to get more of that 80x86 binary today. Prominent among these is IBM, which is now working to perfect its PowerPC 615 chip. Like software emulators, the 615 will let you run 80x86 code on a RISC platform--in this case, a native PowerPC machine. The difference is that the 615 provides a hardware-emulation capability that is significantly faster at executing 80x86 binaries than software-based emulation. How much faster is unknown, and it depends greatly on how IBM approaches hardware emulation.

Approaches to hardware emulation include providing an assist to a software-based emulator, putting a separate 80x86 pr ocessor on the same die as a RISC processor, and creating a hybrid execution pipeline that can handle RISC and CISC instructions. The first solution is the least costly, but it provides the smallest 80x86 performance gain. The second solution provides the best 80x86 performance gain, but it yields very big--and therefore very expensive--chips. The third solution, according to sources, is the one IBM has settled on for the 615.

The 615 has two modes of operation. In one, it acts like a normal PowerPC chip. When the operating system tweaks a specialized mode bit, however, the 615 decodes 80x86 instructions into microcode that is then sent to the execution pipelines. In 80x86 mode, the 615 also tracks things like 80x86 condition codes. Switching between modes is much like a standard context switch.

As for performance, one estimate puts the 615's 80x86 capabilities on a par with those of a 100-MHz Pentium, although there is too little information available to judge this estimate's validity. Also unk nown are the physical characteristics of the chip. These questions will remain unanswered until IBM begins sampling the chip--probably in the first half of 1995. Given that schedule, don't expect to see 615-based systems until late 1995 at the earliest.