Microprocessor Watch

Editor: Michael Slater
Contributors: Linley Gwennap, Keith Diefendorff, Peter Glaskowsky

In This Issue:

• Embedded DRAM Gives Bitboys an Edge
• Profusion Lowers Cost of Eight-Way Servers
• Editorial: A Quiet Revolution at Intel
• Resources: Microprocessor Forum

Embedded DRAM Gives Bitboys an Edge

Bitboys, a Finnish graphics-hardware design firm, has announced a new 3D accelerator it hopes will propel it to the top ranks of 3D-gaming hardware vendors. The Glaze3D 1200 features 9 MBytes of embedded DRAM in four 128-bit banks plus four rendering pipelines in 1.5 million gates of logic on a 130-mm2 die. The chip can fetch 1.2 gigatexels per second from its on-chip texture store and draw up to 600 million pixels per second into its on-chip frame buffer. This pixel-fill rate is nearly twice that of today's fastest 3D chips.

Bitboys (http://www.bitboys.fi) also offers a two-chip set called the Glaze3D 2400 that provides twice the memory and twice the performance of the 1200. The company plans to deliver a five-chip Glaze3D 4800 configuration later next year consisting of four Glaze3D 1200 chips plus "Thor," an unannounced geometry processor.

Bitboys expects to tape out the Glaze3D chip by September. Infineon, its foundry partner, hopes to have first silicon by December. This would be an unusually long fab cycle for a 3D chip, but it is typical of embedded-DRAM processes. Bitboys says it will ship its first products in 1Q00 if first silicon is fully functional. If not, the company faces long delays for each additional spin.

By the time Glaze3D ships, we expect to see next-generation products from current 3D-gaming market leaders 3Dfx and Nvidia. These competing products probably won't match the raw performance of the Glaze3D 2400 chip set, but they will certainly be less expensive to manufacture. If Bitboys can deliver the performance it promises, it's likely to achieve the desired success. --P.N.G

Profusion Lowers Cost of Eight-Way Servers

When we first wrote about Corollary's Profusion chip set (see MPR 9/16/96, p. 9), it was expected to ship in 1997 as an eight-way Pentium Pro chip set. Shortly thereafter, Intel purchased the company, and Corollary has spent the last two years adapting its original design to the faster Pentium III bus interface. Earlier this year, Intel began shipping Profusion board sets (the chips are not available separately) to a number of major server OEMs. Profusion-based systems are already setting records for price/performance on key server metrics, such as the Transaction Processing Council's TPC-C benchmark.

Profusion combines three processor buses, two SDRAM buses, and a programmable cache-coherency mechanism in just two ASICs. Two of the three Pentium III buses each accommodate up to four processors, while the third CPU bus connects the system's four PCI bus bridges.

Two banks of SDRAM, operating independently and concurrently, satisfy the combined main-memory bandwidth demands of eight CPUs and four PCI buses. Up to 32 DIMMs are supported by Corollary's reference design.

Coherent reads and writes that miss in a processor's L2 cache and appear on its bus are tested against the coherency filter for the opposite bank of processors. These filters consist of SRAMs that work like the tags for a direct-mapped L3 cache--but without the data. The filters are used by Profusion to check transactions on one processor bus against the contents of L2 caches on the other bus. A main-memory transaction is initiated in parallel.

Profusion will compete against hierarchical 8-way SMP, switched architectures, CC-NUMA, clustering, and other techniques. Because of its low chip count, Profusion should offer better price/performance than most other MP architectures with up to 16 processors. Intel plans to explore clustering technology to achieve even higher performance on distributed applications.

Given Corollary's excellent track record with PC-based SMP and Intel's ability to create effective products, we expect Profusion to deliver on its promises and become a popular choice for high- end servers. --P.N.G

Editorial: A Quiet Revolution at Intel

Looking at Intel's financial results, it is apparent that the vast majority of the company's revenue and profit comes from PC microprocessors. Yet if you look at the investments, acquisitions, and other announcements Intel has made just since the start of this year, a very different picture emerges. Like the world around it, Intel is moving from PC-centric to Internet- centric, and the sheer number and rate of the new developments is striking.

It will be years, if ever, before PC processor revenues cease to be the major source of Intel's income. But Intel's market share in the PC processor business can't increase dramatically, and its average selling prices are more likely to go down than up. Thus, the core of Intel's business is unlikely to grow much faster than the PC industry itself--and that's not good enough to sustain Intel's ambition.

Some of Intel's investments follow its long-standing approach of backing projects that promise to increase the PC market. An example of just how far Intel will go is the "Intel Play" (toys that are PC peripherals) line it is developing with Mattel. Intel also has pursued a variety of developments to advance digital photography, which makes a superb PC application (see MPR 7/12/99, p. 23); it collaborated with Kodak on Picture CD, for example, aiming to bridge traditional film cameras and digital processing.

Most of Intel's external investment activity, however, is connected to the Internet in one way or another. Intel's increasing efforts in the server market, which have been under way for many years now, are its largest foray. Big servers are a part of the computing business where Intel is weak today, but with the growth of the Web, server farms are sprouting like mushrooms. Some of Intel's investments in this area include Corollary and NCR's server group.

In addition to the billions of dollars of internal investments the company is making in IA-64, earlier this year it founded a $250 million IA-64 venture-capital fund (which includes other investors). The fund's first investments include companies developing database, speech recognition, and CAD software.

Seeking new avenues in which to sell its servers and other equipment, Intel established a program to sell servers and networking equipment to ISPs. And Intel is getting into the ISP business itself: Intel plans to build a series of data centers, with thousands of servers in each, that it will run as "bit factories" for large customers delivering compute- and storage- intensive services via the Web.

Extending its server reach in different ways, Intel spent about $780 million to acquire Dialogic, a manufacturer of software and interface hardware for computer-telephony systems. Intel will provide the servers for these products.

Then there are communications-related expansions of Intel's semiconductor business. Intel's $2.2 billion acquisition of Level One Communications makes it an instant leader in communication ICs for high-speed telecommunications and networking. Intel also acquired Softcom Microsystems, which makes silicon for ATM and SONET equipment.

Intel expanded its networking business with the acquisition of Shiva and recently announced plans to make ADSL modems. Among many other efforts are a collaboration with Hughes to build satellite set-top boxes; with PBS to deliver enhanced digital TV content; and with Proxim to advance wireless home networking. Most recently, Intel invested $50 million in Pacific Century CyberWorks to deploy high-speed Internet services to Intel-based set-top boxes in Asia.

Even Intel's processor lineup is changing as a result of the new focus. In the case of the x86 line, Pentium III is being heavily pushed as enhancing the Internet experience--despite a paucity of substance behind the claim. More significant, Intel's nearly accidental acquisition of StrongArm gave the company an outstanding processor with which to pursue communications equipment. Going a step further, Intel is developing a "network processor," which remains only vaguely defined but will be aimed at routers and other infrastructure equipment. And the company entered into a joint effort with Analog Devices to create a new DSP architecture, rounding out the spectrum of processors it might need for virtually any communications-related application.

Although Intel's preeminence today is based on the overwhelming success of a single microprocessor architecture, the company's vision for the future is much broader. Intel's server and communications investments may provide the engine for growth, should the PC market falter--and they make the company less vulnerable to a collapse in pricing for PC microprocessors.

As Intel moves in so many directions, the company is spreading itself so thin that it will surely blunder in some areas. It will also face many competitors in each new area, and rarely will it have the advantage of a powerful architectural franchise. But Intel has wisely, yet quietly, stopped its rallying cry of "the PC is it," while it quickly prepares for a new era. --M.S.

See http://www.MDRonline.com/slater/newintel for more on this subject. I welcome feedback at mslater@mdr.cahners.com

Resources: Microprocessor Forum

Microprocessor Forum, October 4-8 in San Jose, is filling up fast. Reserve your place today to ensure you don't miss out on the many disclosure to be made there, including details on Intel's Coppermine and Merced microarchitectures; upcoming high- end Alpha, SPARC, and PowerPC processors; the first Athlon processor for servers; Sun's first MAJC chip; a radical new chip from startup Cradle Technologies; a PC graphics chip from newcomer ArtX; National's first information appliance on a chip; a new 64-bit MIPS core for embedded applications; Hitachi's ST-5; three new DSPs; and much more.

Check out the full agenda today at http://www.MDRonline.com/mpf and then register on-line or call 800.700.4004 or 707.824.4004.