| Don't be fooled by fast CPUs: The PC of 1995 is little
more than a souped-up IBM AT from 1984, which was a minor
improvement over the IBM PC of 1981, which was based on
technology from the 1970s that wasn't so hot to begin
with. Fortunately, that's all about to change. Some differences will be externally obvious. Think sleek shapes in black or Caribbean colors. Imagine a collection of compact components and a svelte LCD. Reliable voice recognition is on its way. Cables will give way to infrared beams. And built-in video cameras will let your PC stare back at you. But to really appreciate the new PC, look under the hood. New technologies will unleash the full potential of modern operating systems and microprocessors, and performance bottlenecks from albatrosses like the ISA bus will disappear. Four of these technologies are available now: Plug and Play (PnP), universal serial bus (USB), unified memory architecture (UMA), and native signal processing (NSP). Together, these technologies will create a more integrated PC that is easier to configure, has significantly faster I/O, uses memory more efficiently, offers more features without extra hardware, and costs relatively less than today's equivalent systems. The new PC won't make its debut with fanfare and fireworks. Rather, it will evolve steadily over the next five years (see the illustrations in the sidebar "On the Cover"). Here's what to expect during that evolution. Still Praying for Plug and Play To see tomorrow's PC, look at today's Macintosh. Macs are far from perfect, but they have pioneered many technologies now coming to PCs. Plug and play has been standard since the 1980s, along with reasonably fast and easy-to-use I/O interfaces such as SCSI and the Apple Desktop Bus (ADB). Numerous Macs employ a unified frame buffer/memory architecture, and Power Macs use NSP for some telephony applications. Moreover, the latest Power Macs are the first mainstream computers to discard their legacy I/O bus (in this case, NuBus) in favor of PCI. Playing catch-up with Apple, however, is not what PC designers have in mind. The goal is to leapfrog the Mac by adopting superior versions of these technologies whenever possible. Leading the charge are Intel and Microsoft, who may be the only companies powerful enough to direct wholesale changes in the industry. (See the sidebar "Intel and Microsoft: The Agents of Change".) PnP is high on Microsoft's agenda because it's probably the most glaring remaining difference between a PC and a Mac. Although Windows 95 is loaded with PnP features (the Device Manager, dynamic configuration, the Add New Hardware Wizard, and more), they aren't much help if the computer and peripherals don't support PnP as well. Without a PnP BIOS and PnP devices that cooperate with each other and with the OS, painless system expansion will remain just slightly out of reach (see "Transforming the PC: Plug and Play," September 1994 BYTE). Fortunately for long-suffering PC users, there's strong movement toward PnP. All signs point to wide adoption of PnP throughout the industry in 1996, even if the only motivation for some vendors is to control their skyrocketing technical-support costs. All along, Apple has preached the gospel that hardware and software must be tightly integrated. Unfortunately for Apple, it's hard to sell this integration as a feature precisely because it's so transparent. It was easy for Apple to achieve this level of integration because, until recently, the Mac was a closed, proprietary system available only from Apple. Ironically, as the PC becomes more like a Mac, a Mac may become more like a PC, since Apple now is licensing the Mac OS to clone makers and porting it to the PowerPC common hardware platform. Port Power USB represents another major improvement in PC architecture. This external I/O interface will begin appearing in commercial desktop systems next year. If it succeeds, USB could eventually replace a whole slew of ports on today's PCs. This won't happen immediately, because the huge installed base of PCs impedes rapid change. Nevertheless, I/O bottlenecks will push PC vendors to add an interface designed with 1990s technology. Venerable RS-232 serial ports offer speeds of from 9.6 to 115.2 Kbps, depending on the type of universal asynchronous receiver/transmitter (UART) in the computer. USB's maximum bandwidth is 12 Mbps — although actual data throughput is more like 8 Mbps — including a 1-Mbps subchannel for low-speed devices like a mouse or keyboard. That's enough bandwidth to handle everything from keyboards and mice to video monitors, modems, scanners, printers, ISDN adapters, and MPEG-2 compressed video. In addition, USB is both asynchronous and isochronous. Isochronous transfers, such as audio and video, get top priority, assuring that time-sensitive data streams are not interrupted. USB lets you daisy-chain up to 127 devices in a tiered-star topology: Each device can house a USB hub to which additional devices can connect. Cable segments can be 5 meters long. It supports hot-plugging, so you can add or remove devices without powering down the computer. If the devices and the OS support PnP, the appropriate device drivers can automatically load and unload. USB is also cheap to implement. Some of Intel's PCI chip sets will soon include USB logic, so the only additional cost is for the 35-cent external connector. As other chip makers incorporate USB into their products, the new serial bus will become a standard feature on new PCs. Inexpensive USB might also encourage engineers to include USB hub logic in external peripherals, which would provide for extra USB ports for daisy-chaining additional devices. Cables have only four wires, allowing compact connectors roughly as wide as a staple. Small size is important for notebooks, palmtops, and downsized desktop PCs. For all these reasons, USB appears to be edging out alternative I/O interfaces that have been proposed as the next-generation standard, including GeoPort, Access.bus, FireWire, and SCSI. GeoPort is a slightly enhanced version of the 12-year-old RS-422 serial ports on Macs. Apple already equips Power Macs with GeoPorts to provide telephony functions. For instance, the GeoPort Telecom Adapter connects a phone line to the Mac while the PowerPC processor emulates a fax modem. Apple formed a consortium called Versit with IBM, AT&T, and Siemens to promote GeoPort as an industry standard, and it will be available for PCs in the form of PCI and PC Cards (formerly PCMCIA). However, GeoPort's relatively slow throughput of 2 Mbps probably eliminates it from contention as the next-generation external I/O port for PCs. Access.bus is even slower than GeoPort — its data rate is only 100 Kbps. Invented by Philips, it was never intended to be a general-purpose high-speed interface. Instead, it was designed for low-speed peripherals, such as keyboards and pointing devices. It's also used for the Video Electronic Standards Association (VESA) Display Data Channel (DDC), a control interface for video monitors that Microsoft recommends for future PCs. But even though Access.bus costs less to implement in peripherals and on motherboards, its inability to support a wide range of peripherals puts it at a disadvantage with USB. FireWire — officially known as IEEE-P1394 — is backed mainly by Apple, with chip support from Texas Instruments. It has significant advantages over USB. FireWire's data rate is 100 Mbps and will soon be extended to 200 Mbps and 400 Mbps. Also, FireWire is being promoted as a standard interface for consumer video devices, such as digital VCRs and cameras. If widely adopted, it could accommodate much faster peripherals than USB and provide the crucial digital link between personal computers and future consumer electronics products. But because FireWire logic isn't inherent in chip sets and peripherals, it costs more to implement than USB. Apple has yet to build FireWire into any of its Macs, and it's even less likely to be integrated in PCs. It will probably be available on a PCI board for those who want it in 1996. SCSI remains a contender. When Apple standardized on SCSI in 1986, it spawned a healthy market for SCSI-based scanners, hard disks, CD-ROM drives, and other peripherals. When PC users began upgrading their systems with CD-ROM drives, SCSI adapters zoomed in popularity. Today, SCSI is an important cross-platform interface standard that's also found on some Unix workstations. SCSI-2 introduced faster throughput in two flavors. Fast SCSI-2 doubled original SCSI speeds to 10 Mbps on an 8-bit bus. Wide SCSI-2 can use a 16- or 32-bit bus for throughput of 20 Mbps. Fast and Wide SCSI achieves speeds of 40 Mbps. But SCSI has several problems. Current implementations support only eight devices per chain, far fewer than rival serial interfaces. Each end of the daisy chain must be terminated, a prime cause of configuration woes. Each device requires a unique ID number, and some devices don't support all IDs. Because SCSI is a parallel interface, the connectors are relatively large, the cables are thick and expensive, and longer cable lengths are unreliable. SCSI doesn't allow hot-plugging, either. Work is progressing on new versions of SCSI to address these flaws. The proposed SCSI-3 standard calls for more than eight devices per chain, automatic ID assignment, and faster throughput. There are even proposed serial versions of SCSI-3 that could use fiber-optic cables or gallium arsenide technology to attain speeds as high as 1 Gbps. So SCSI still has a bright future as the high-speed alternative in applications that need peripherals that would quickly saturate the capacity of USB. RAM Reunification Since the original IBM PC of 1981, almost all PCs have maintained separate memory for their frame buffer, the block of memory in which the screen image is mapped. This wasn't a problem in the days of monochrome character-mapped video because the frame buffer could be as small as 2 KB. But modern GUIs have created demand for high-resolution, true-color bit-mapped screens. A 640- by 480-pixel screen with 8-bit color requires a frame buffer of 300 KB; 1024 by 768 pixels with 24-bit color requires 2.25 MB. It's not just an issue of more memory. A dedicated frame buffer is fixed in size — no matter which screen mode you're using, the frame buffer always contains enough memory to accommodate the highest resolution and color depth it supports. Your software can't use the leftover memory. Because of the way DRAM is packaged and configured in graphics subsystems, megabytes of memory can go to waste. UMA unifies the frame buffer with main memory. By allocating just enough RAM to handle the current screen mode, UMA frees up memory for other purposes. Eliminating the dedicated frame buffer is expected to trim about $50 off the cost of a typical system. Chip sets that support UMA are coming from Weitek, Opti, VLSI, Chips & Technologies, Intel, and Cirrus Logic. PCs with UMA should appear this year or early in 1996. In all likelihood, UMA will come first to low-end PCs. High-end systems will probably avoid it, because UMA sacrifices performance when the CPU and the graphics controller try to access main memory at the same time. Cirrus Logic estimates the performance hit is about 5 percent for a 16-MB system with a secondary CPU cache, and 10 percent to 15 percent for cacheless systems. An 8-MB system would suffer more. Reduced performance is why Apple, which has used a similar scheme for years, still retains dedicated frame buffers in some Macs. Mixed Signals Still up in the air is how PCs will handle signal-processing tasks, such as audio, video, and telephony. This question pits Intel against some other chip makers and against Microsoft. PCs have taken advantage of general-purpose digital signal processors (DSPs) for years. In 1992, Atari introduced the Falcon030, the first personal computer with a general-purpose DSP. Then Apple built AT&T's 3210 DSP into the Quadra 660AV and 840AV Macs. IBM and Texas Instruments announced the low-cost MWave DSP for PCs. And Microsoft said Windows 95 would have a DSP Resource Manager Interface (RMI), an API layer designed to insulate programmers from DSP-specific features. But the trend was derailed. First Apple, then Intel, claimed its latest CPUs could handle many DSP tasks natively. Apple's GeoPort Telecom Adapter eliminates the need for a modem by connecting the Mac directly to a phone line while using the CPU to emulate a modem in software. Intel is carrying the concept even further by defining an NSP reference platform for PC vendors and extending the functions to include wave-table audio and software-only video playback. Instead of handing off those functions to dedicated chips and special hardware, Intel's NSP approach uses the CPU to perform audio and video tasks. Intel has demonstrated a 90-MHz Pentium system that uses NSP to display full-motion video in two separate windows while playing eight independent audio tracks. Partly for these reasons, Microsoft dropped RMI out of Windows 95 and turned it over to Spectron Microsystems, which makes SPOX, a real-time DSP kernel. The lack of a DSP API built into the industry's leading OS is a setback to the widespread integration of DSPs. In the view of Compaq and other systems vendors, NSP is a cost-effective technology for delivering speech, audio, and other multimedia functions to the business market. But not everyone is ready to shelve DSPs. "The functions that NSP can provide by using the host [CPU], a few dollars' worth of hardware can provide," says Raphael Mehrbians, a senior product manager at Cirrus Logic. Also, some of that extra hardware is needed anyway, even with NSP. For backward compatibility with DOS programs, PCs need a Yamaha OPL-3 sound chip. The Pentium's wave-table sounds won't be heard unless there's a standard audio codec chip, such as a Crystal Semiconductor 4232. And even though a Pentium can handle such modem functions as data compression and communications protocols, DSP promoters like TI question whether the best use for a high-end CPU is to emulate the modem's data pump. So general-purpose DSPs may yet appear on PC motherboards, or perhaps be integrated in future CPUs, just as FPUs were. Intel is known to be developing a "multimedia Pentium" (code-named the P55C) for release in 1996. It is expected to add new instructions for signal processing, but at press time details were not available. It may have a multiply-accumulate function or special instructions for decoding MPEG video — a feature introduced last year in the UltraSPARC processor from Sun Microsystems (Mountain View, CA). False starts and growth spurts will buffet the PC industry as it juggles the conflicting issues of cost, performance, and compatibility. Yet change is essential as the competition for desktops becomes fiercer. By mid-1996, expect systems that support the PowerPC common hardware platform, which is backed by IBM, Motorola, and Apple, and designed with the benefit of hindsight. When PCs were primarily used for word processors, spreadsheets, and databases, the original IBM system architecture was adequate. But multimedia and richer data types have exposed shortcomings in PCs. For the last few years, the industry has been patching the old architecture to keep it competitive. Now the architecture is finally getting the overhaul it really needs. Sidebars:
Where to Find
|