The gradual movement by companies--and vendors--from RISC to EPIC is part of a long evolution of processor technologies.
In computing's early days, software interfaced directly with a computer's processor. Programming was an arcane science, but the code ran relatively fast. By the 1960s, high-level, compiled languages such as Cobol and Fortran made software easier to write, though performance was less efficient. Mainframes, minicomputers, and the PCs that emerged in the late '70s and early '80s contained processors that used a design called Complex Instruction Set Computing. One well-known example: Intel's x86 design.
In the '80s, Reduced Instruction Set Computing--which grew out of research at IBM, Stanford University, and the University of California at Berkeley--brought systems that consumed more instructions per CPU cycle than CISC machines, gaining speed by making each instruction simpler. Since there was no dominant semiconductor supplier then, systems vendors designed their own RISC chips, plus the memory, input/ output systems, software libraries, and operating systems for each new computer. The integration led to some great designs, such as IBM's RS/6000 and Digital Equipment Corp.'s Alpha Server. But closed systems have their price. Digital saw its influence decline as high-powered Unix workstations from vendors such as Sun Microsystems emerged, partly because there wasn't enough third-party software available on Digital's VMS operating system.
Itanium's architecture, co-designed by HP, uses an approach called Explicitly Parallel Instruction Set Computing. The design offloads more complexity to software compilers, which decide which instructions in each word, or string of bits, can be executed simultaneously. Itanium chips also contain more registers on each processor to store numbers and enough headroom to process all possible branches of a calculation in advance, which keeps the CPU from having to wait for data.
The first Itanium chip, released last year, disappointed users. Itanium 2 is faring better. It's second only to IBM's Power 4 chip for integer processing in a single-CPU system, according to the independent SPEC CPU2000 benchmark test, and faster than Sun's UltraSparc III. For floating-point operation performance, Itanium 2 tops all major RISC chips. "They beat UltraSparc, and that's a head-turner," says David Freund, an analyst at market researcher Illuminata.
But migrations to Itanium could be slow. Market researcher Gartner Dataquest predicts Itanium-based systems will account for $3.1 billion in worldwide server sales by 2007 or about 6% of the total server market. "That's not a lot," analyst Jim Cassell says. One factor that could hinder acceptance: Intel keeps improving the performance of its 32-bit Pentium and Xeon chips.
Intel also is reportedly hedging its bets by developing a set of 64-bit extensions, code-named "Yamhill," which could be turned on in future Pentium chips. Using the Yamhill code, users could run 32-and 64-bit Windows and Linux apps on an x86 chip, without recompiling them or paying a performance penalty.
Advanced Micro Devices Inc. is taking a similar approach with its Hammer line of chips, due next year. Intel won't comment on Yamhill.