Processors begin 64-bit push
By Rick Merritt
EE Times
October 14, 2002 (2:16 p.m. EST)
SAN JOSE, Calif. — Microprocessors will make a fresh push toward 64 bits this week, both on the desktop and in embedded systems. IBM Corp. will detail at the Microprocessor Forum here its first 64-bit PowerPC microprocessor, a 1.8-GHz CPU that's expected to power next-generation Macintosh computers. Analysts expect Apple could roll out 64-bit Macs starting late next year, turning up the heat on Intel Corp. to extend its flagship Pentium processor to 64 bits.
On the embedded side, NEC Electronics this week will roll out a 64-bit MIPS processor that's positioned as a system-on-chip core. And Tensilica Inc. will describe an architecture that incorporates user-defined 64-bit instructions for better parallelism, though the company is keeping the 32-bit core architecture intact.
Essentially a derivative of the company's Power4 microprocessor, IBM's PowerPC 970 adds 64-bit PowerPC compatibility, an implementation of the Altivec multimedia instruction-set extensions and a fast processor bus supporting up to 16-way symmetric multiprocessing.
The CPU is expected to spark debate over whether the desktop is ready to break through the 4-Gbyte addressing of 32-bit architectures. Whatever the answer, the 970 would give Apple Computer Inc. a chance to deliver high-performance 32-bit systems that could later be upgraded to full 64-bit computers. That's something Advanced Micro Devices Inc. will offer PC makers with its upcoming Opteron processor but that Intel cannot deliver with the current Pentium road map.
Apple declined to comment last week, but a source close to the company said the computer maker indeed will use the 970. The device could also show up in communications gear and in entry-level servers from IBM and other OEMs.
"Apple would have to be crazy not to use this part," said Peter Glaskowsky, editor-in-chief of the Microprocessor Report, which hosts the annual Microprocessor Forum. "Its performance will be in the upper reaches of any CPU. Apple would be able to produce for the first time machines that not only have great performance but support full 64-bit addressing."
With AMD and Apple both moving toward 64-bit desktops, Intel will be forced to respond, said Kevin Krewell, senior analyst at the Microprocessor Report. "I think it's a no-brainer for Intel to extend its Pentium line beyond the current 32-bit architecture. We'll see it first in the Xeon because the servers and workstations need the extra data-crunching capabilities most. My best guess is they will do it in 2004 or 2005," Krewell said.
Intel's Pentium-based Xeon processors already have a 36-bit page-mode addressing scheme that's supported in Windows XP and many Linux distributions. However, the mode is "awkward and limited," Krewell said.
"We see it as quite a long while before 64-bit will bring benefits to end users at the desktop," said an Intel spokesman, responding to the IBM news.
Nathan Brookwood, analyst for Insight64 (Saratoga, Calif.), agreed, noting that 64-bit is best suited for servers, not desktops. "Most PCs have yet to bump up against the 4-Gbyte memory-addressing limit of 32-bit chips," he said. "A 64-bit Macintosh will deliver only marginal benefits until Macs need 4 Gbytes of RAM. But I don't doubt Apple could do a great job of marketing the feature if they choose to adopt it."
"Is the desktop crying out for 64-bit? Maybe not," said Peter Sandon, manager of PowerPC architecture at IBM Microelectronics. "But maybe that's because it's just not been available or it's been seen as expensive. I certainly think we are implementing it the way it should be done."
The real question is not whether 64-bit is coming to the desktop, but "how much of that capability will really be used," said Dean McCarron, principal of Mercury Research (Phoenix). "Apple's in a better position than AMD to use it aggressively since they control their own OS."
Indeed, both IBM and AMD are taking advantage of Intel's decision to create separate 32-bit and 64-bit architectures with its Pentium and Itanium lines, leaving a strategic hole for migrating desktops to 64 bits. The flip side is that Intel is executing very well on its Pentium 4 line, with plans to release by the end of this year a 3-GHz version with the company's homegrown simultaneous-multithreading technology, called Hyperthreading.
"Intel has a significant clock speed advantage over IBM, but as we've seen with AMD it's not all about clock speed," said Krewell of the Microprocessor Report. "The PowerPC instruction set is more efficient than the Pentium's, making a direct comparison with the Pentium 4 difficult."
In terms of die size, a rough measure of cost, the PowerPC 970 measures 118 mm2, against 131 mm2 for the Northwood 2.X-GHz Pentium 4. Both the IBM and Intel parts are being made in 130-nanometer CMOS on 300-mm wafers.
Analyst Brookwood noted that the 970 represents a return to the Mac market for IBM after a two-year lull. Motorola Inc. has been supplying most of Apple's PowerPC needs except for CPUs for some G3 notebooks which are coming from IBM, he said.
How IBM did it
IBM's approach to implementing a 32/64-bit architecture appears straightforward. The 970 supports full 64-bit registers and addressing. When a flag bit is sent it triggers a 32-bit mode in which the high-order words on an arithmetic logic unit and on memory addresses are ignored. In either 64- or 32-bit mode, the processor issues up to eight instructions per clock cycle.
Thus, 32-bit PowerPC applications run unchanged on the 970. However, a 32-bit operating system would have to support new data structures in its memory management unit and new interrupt handlers.
"There's not too much more to it," said IBM's Sandon. "Certainly we found it a compelling idea that there is market segment asking for 64-bit systems and this allows them to still make use of 32-bit code."
By contrast, AMD uses a 64-bit prefix in front of X86 instructions to tag a 64-bit operation. The technique requires referring to separate 64-bit registers and extended memory addresses. That method is "very elegant in terms of not disrupting the X86 architecture but may not be able to support mixed 32- and 64-bit operations as well as IBM's approach," said McCarron of Mercury Research.
Besides the 64-bit capability, the PowerPC 970 includes IBM's first support for what Motorola calls the Altivec instruction set and Apple refers to as its Velocity engine. IBM simply cites the capability of the 160 vector instructions in its SIMD engine to speed graphics and multimedia operations. IBM had opted not to support the instruction extensions in its 32-bit PowerPC family but will bring them into the new 64-bit line. "Adobe has been the premier company taking advantage of [the Altivec extensions]," said Krewell of Microprocessor Report.
The 970 also sports a cache-coherent, 900-MHz processor bus capable of data rates up to 6.4 Gbytes/second. It will support symmetric-multiprocessing configurations of up to 16 CPUs. That capability could be valuable for Apple's new line of Xserve entry-level servers.
Unlike the original Power4, from which the CPU was derived, the 970 supports only one internal processor core. Sandon would not comment on plans for future members of the new 64-bit PowerPC family.
Embedded angle
The IBM CPU is one of about 18 new computer, embedded or network processors to be detailed at the Microprocessor Forum. "I expect there will be a fair amount of discussion about this part," said Glaskowsky. But the embedded side of the fence, especially the Tensilica and NEC devices, will also attract its share of attention this week.
NEC Electronics intends its Sapphire processor as a system-on-chip core. Based on the MIPS instruction set, the 0.13-micron Sapphire will run as high as 800 MHz. NEC has also started work on a 90-nm version of the core that will reach speeds of 1 GHz, said Arnold Estep, senior marketing manager at NEC Electronics (Santa Clara, Calif.).
To home in on storage, networking and encryption applications that can benefit from the larger addressing of 64-bit, NEC is working with the Linux community, Wind River Systems and Microsoft on 64-bit versions of their respective Linux, VxWorks and .Net operating systems, Estep said.
But because of the large memory footprints needed to cover the bigger addresses, the demand for a 64-bit embedded operating system has been limited, observers said.
Jim Liband, director of embedded-platforms marketing at Wind River, said his company sees no immediate need to introduce a 64-bit OS for the embedded space since only about 1 percent of its customers are asking for one. "They want an OS that runs on [64-bit] parts but it doesn't have to be for full-fledged 64-bit software," Liband said.
Among embedded processors, Tensilica will discuss a new VLIW platform that will become a base architecture the company will offer. And Motorola will detail version 5 of its ColdFire embedded 32-bit processor.
Among other news at the MPU Forum, startup MemoryLogix will announce a Pentium II-class synthesizable X86 core for embedded applications. The part will be comparable in die size and cost to current 486 parts, said Glaskowsky, and will be capable of running at several hundred megahertz. Users will be able to customize the chip's cache size and whether it uses a floating-point processor, he added.
AMCC, Intel and Cisco will discuss separate 10-Gbit/s network processors at the conference. The parts are all enhancements to existing 10-Gbit/s NPUs, though Intel's offering does have "one feature I think a lot of people will find surprising," Glaskowsky said.
The T3 is Cisco Systems Inc.'s third-generation Toaster network processor, designed for use in its own systems. "Seeing what Cisco has done and how fairly straightforward their device is should be interesting — especially since it means this company will not be a customer for some other vendors," said Glaskowsky.
For its part, ARM Ltd. will discuss the first two ARM 11 processors. Running at 400 MHz and up, the cores "feature an eight-stage pipeline, a new memory system, a vector floating-point unit and a quad 64-bit Amba AHB-lite bus interface," according to the conference agenda.
"These processors sort of stretch how much performance you might want to have in a cell phone, but they should find good use in PDAs," said Glaskowsky. |
