And Chip revenue doesn't grow because of PCs.....................
techweb.com
July 05, 1999, Issue: 1068
Section: Analysts' Outlook
--------------------------------------------------------------------------------
Designs Drive New MPU Choices
Michael Slater
As the microprocessor industry heads into the new millennium, it is facing more opportunity and greater change than at any time in its nearly 30-year history. The changes are the result of a combination of forces: ever-increasing transistor counts, the dynamics of the personal computer industry, the explosive growth of the Internet and the digitization of all media.
Like all semiconductors, the key driving force behind the evolution of microprocessors is the relentless doubling of transistor counts approximately every 18 months. Unlike a memory chip, however, whose function remains constant even as the size increases-and it takes no great innovation to figure out what to do with twice as many transistors-it isn't always obvious how to use increased transistor budgets in a microprocessor.
Although many common PC applications don't demand leading-edge performance, some do: notably 3-D (including games) and image processing, which is moving into the mainstream with digital photography. More buyers are deciding that low-end systems are good enough, but that doesn't mean the end of the road in performance demand-just a lag between the hardware capability and the emergence of software with wider appeal that can exploit it. As software with 3-D user interfaces, natural-language understanding and intelligent agents becomes widespread, demand for processor speed will increase.
Because of the constraints of software compatibility and the dynamics of the market, innovation in PC processors is restricted in some ways. Changing the base instruction set isn't acceptable, since users show little interest in PCs that won't run all the existing software. New applications that process images, video and audio make stringent demands on processor performance, however. Fortunately, those applications can benefit greatly from single-instruction, multiple-data (SIMD) instructions that enable multiple data points to be processed in parallel. This is a far more efficient approach, for those applications that can benefit from it, than trying to build very wide superscalar machines.
Intel's MMX added integer SIMD to the X86 architecture; more recently, AMD's 3DNow and Intel's SSE added floating-point SIMD. Such extensions have allowed the venerable X86 architecture to keep up with changing work loads and deliver impressive performance despite the handicaps of the original architecture.
Market dynamics have worked curiously against aggressively superscalar designs. Processors that do lots of work in each clock cycle tend to be slower, and most PC buyers remain focused on speed, rather than benchmark performance. This has created a disincentive to give up any clock speed in pursuit of parallelism. SIMD instructions are a perfect fit, since they provide parallelism without compromising clock speed.
PCs will continue to drive volume demand for high-performance processors. But as digital media proliferate, so will the number of devices that access them. Audio was the first to go digital, with the shift to CDs. High-end video has gone digital, in the form of DV camcorders and DVD players, and those technologies will move into the mainstream in the next few years. The Internet is rapidly converting mail, news and many types of commerce and publishing to digital form. Handheld organizers are replacing paper-based organizers. In this new world, the PC will play an important role, but it will not be the only access device.
Information appliances open up opportunities for all architectures in applications that have formerly been the nearly sole domain of the PC and its X86, and the Macintosh with its PowerPC processors. When building a general-purpose digital device (i.e., a personal computer), software compatibility is the overwhelming factor in microprocessor selection. But when creating a device with a focused purpose, such as Web browsing and e-mail, the range of software required is much narrower and the processor can be chosen based on price, performance, power consumption and other relevant characteristics. That allows for a much more competitive landscape. The important standards for that kind of application are not instruction-set compatibility but communication protocol and data format compatibility. One of the Web's contributions is that is has codified a relatively simple set of protocol and data standards (as compared with those in the PC realm); unlike the PC world, it isn't necessary to implement a vast and ever-shifting set of application programming interfaces to provide good compatibility.
Information appliances typically do not demand the very fastest microprocessors. That fact enables the use of relatively small processor cores, leaving room to integrate many other functions on the same chip as the processor. That trend will accelerate with the emergence of 0.18-micron technology and will become overwhelming with even denser processes. At the same time, it is more practical to integrate all the functions of an information appliance on one chip than it is for a PC, since the functions are more limited in scope and user expansion is not as much of an issue. Furthermore, low cost and low power are very important, making high integration valuable.
System-level integration will occur in PCs as well, but it will be slower to arrive and more limited in scope. Past attempts at integrating system functions with PC microprocessors, such as Intel's SL family and Cyrix's MediaGX, have had only transient success. If PC hardware continues to evolve as fast as it has in the past, integration will be challenging. PC makers like to be able to take one system board and add a low-cost graphics board for a value system and a high-end board for a gaming system; integration makes that difficult, if not impossible. Nevertheless, integration ultimately will win out; its cost benefits will be compelling as process technology advances, and many users will tire of the technology treadmill and be willing to accept more constrained designs in return for lower prices and higher reliability.
-Michael Slater (michael@mslater.com) is principal analyst at Cahners Microdesign Resources (www.mdronline.com).
Copyright (c) 1999 CMP Media Inc.
|
