Inside PC Processors

Socket 7 vs. Slot 1

From a system hardware perspective, Cyrix and AMD are pursuing a strategy fundamentally different from Intel's. The Pentium MMX/233 is probably the end of the Pentium interface, called Socket 7, for Intel. For continued performance and technology increases, Intel is shifting to the Pentium II interface, called Slot 1.

AMD and Cyrix have stuck with Socket 7 for several reasons. First, the Slot 1 bus is complex, and details of its operation weren't publicly available when the design of the competing processors began. Second, Intel is believed to have system-level patents that could be asserted against system makers that used non-Intel Slot 1 processors. In AMD's case, there are rumored to be specific exclusions in the company's patent cross-license with Intel that limit what AMD can do with Slot 1. And fundamentally, Slot 1 simply isn't needed today for mainstream desktop systems.

Slot 1 uses an entirely different system bus from that of Socket 7. The Slot 1 bus, functionally identical to the Pentium Pro bus, is a highly pipelined design that allows many transactions to proceed in parallel. It also allows bus operations to finish in an order different from that in which they began. This design makes it much more effective for multiple-processor systems than the Socket 7 bus.

In a single-processor system--in other words, the vast majority of desktop systems--the Slot 1 bus advantages are minor. The Slot 1 design's big advantage in such systems is that there is a second bus, hidden inside the cartridge, that connects the processor to the cache SRAM. This approach, which Intel calls the Dual Independent Bus (DIB) architecture, enables the cache interface to run faster than the system bus. It also makes it possible for Intel to use different cache interface speeds while keeping the system interface constant. At 266 MHz, the Pentium II L2 cache bus runs at 133 MHz--twice the speed of the 66-MHz system bus in a Pentium system. In addition, all of the system bus bandwidth is available for main memory and input/output traffic, whereas a Socket 7 system must handle L2 cache transactions on the system bus as well.

To mitigate the limitations of the Socket 7 bus bandwidth, AMD and Cyrix each gave their latest processors twice the L1 cache of Intel's Pentium MMX or Pentium II. This means that more memory accesses are satisfied by the on-chip cache, so there is less traffic on the system bus and its speed limits are less constraining. The fact that the L2 cache must run at the system bus speed, however, even as CPU speeds approach 300 MHz, is a serious limitation.

One way to get around this limit is to increase the system bus speed. Cyrix already uses a 75MHz bus in its 6x86MX-PR233, and both AMD and Cyrix will move to 83 or 100 MHz bus speeds for some future CPUs. This pushes the limits of what can be achieved with the Socket 7 bus, and chip set design at these speeds will be difficult. Note that even at 100 MHz, the system bus would be slower than the L2 cache bus in the slowest Pentium II system. Intel plans to move to a 100-MHz system bus for the Pentium II in 1998, and the Pentium II's L2 cache speeds will exceed 150 MHz.

The midterm solution for AMD, Cyrix, and other x86 contenders is to integrate the L2 cache on the processor chip. In 1998, at least two processors are likely to be available that put a 256K L2 cache on-chip. This cache will have its own private bus to the processor--just like a Pentium II cartridge--and everything will be on a single piece of silicon. This technique should keep Socket 7 systems viable, except at the high end of the market, through 1998.

In time, however, Intel's competitors will have to move to a new bus design. They might produce CPUs compatible with Slot 1 (or its successor) in 1999, or they may move to a completely new bus design.

AGP Adds a Twist

The biggest change in PC system architecture this year will be the arrival of the Advanced Graphics Port. AGP provides a separate, dedicated connection for the graphics controller, enabling the processor to send commands to the graphics chip faster and allowing the graphics controller to transfer data from main memory at a much higher rate. This latter feature will make it more practical to store memory-intensive 3-D texture maps in main memory, instead of requiring additional memory as part of the graphics subsystem. AGP in essence is an enhanced version of PCI that supports much higher transfer speeds. It allows only a single device to be connected, however, so it supplements, and does not replace, PCI.

To promote the transition to Slot 1, Intel does not plan to support AGP in its Pentium chip sets or motherboards. This presents Cyrix and AMD with a new challenge: Today, Intel chip sets predominate--even in systems using Cyrix and AMD CPUs. To move into the AGP era, system designers will have to depend on non-Intel chip sets to accompany non-Intel CPUs.

Several chip set makers have committed to producing AGP chip sets for Socket 7 processors. Implementing AGP is complex, however, and it remains to be seen whether these chip sets will deliver performance comparable to Intel's LX chip set for Pentium II systems. To help ensure that competitive AGP chips sets are available for the K6, AMD has gone into the chip set business itself. AGP chip sets from Intel and others are due this fall.

3-D and Multimedia Acceleration

Many of today's PC applications don't require greater processor performance; they run perfectly well on a Pentium/133. Many of the most demanding applications are those that use 3-D graphics or video, and these applications can be enhanced--within limits--by hardware accelerators.

For 3-D graphics, functions can be divided between the processor and the graphics accelerator in many ways. With a simple, 2-D-oriented graphics controller, all the 3-D functions are handled in software; part of Microsoft's Direct3D software is an emulation library that performs all the required functions, using MMX if available. With most of today's 3-D accelerators, the rendering function--drawing filled triangles on the screen--is handled in the accelerator, while the geometry function--manipulating the 3-D image's mathematical model to produce a set of 2-D triangles--is done by the host CPU. Geometry processing generally requires floating-point computations, so CPUs with good FP performance do best on this task.

In our tests, the benefits of MMX show up clearly when performing 3-D rendering in software. Once a good 3-D accelerator is added, though, the usefulness of MMX for 3-D evaporates, so the Pentium Pro, for example, provides nearly the same performance as the Pentium II. The slow floating-point execution of the Cyrix and AMD parts results in lower 3-D performance even when a good 3-D accelerator is used, however.

To make the most of your money, consider buying a slightly slower CPU and spending the savings on MPEG or 3-D accelerators. These accelerators won't help performance on applications that don't use these functions, of course, but they can provide a much bigger boost in 3-D and video performance than you would get from moving up one speed grade in a processor line. A $100 CPU with $500 in graphics hardware could yield a better multimedia system than a $500 processor with $100 in graphics hardware--though this is not, as you may imagine, the way the processor vendors want you to think.

Be careful not to mate a fast host CPU with a slow 3-D accelerator. With a Pentium II, for example, the mainstream 3-D card used in the tests actually acted as a "decelerator"--it delivered a lower 3-D WinMark score when its acceleration functions were enabled. This is because the Pentium II, using the Direct3D emulation code with MMX, was able to render faster than the hardware. The high-end 3-D card did provide a performance boost, however, even with the fastest Pentium II.

The most advanced 3-D accelerators have fast FP engines to handle geometry processing as well as rendering. Today these chips are used primarily in CAD applications, generally using the OpenGL API. Geometry acceleration isn't supported in the way most games are written, using Direct3D or other game-oriented APIs. In the next year or so, lower-cost 3-D chips with geometry acceleration will appear, and the Direct3D API will evolve to support such accelerators more readily. Once these technologies are widely available and games are modified to take advantage of them, the FP performance of the host CPU may become unimportant for 3-D. A lot of software needs to change before this will be the case, however. Furthermore, if the 3-D chip takes over geometry processing but doesn't have faster FP execution than the host CPU, the "deceleration" phenomenon seen with the Pentium II and today's mainstream 3-D accelerators could recur.

From the September 23, 1997 issue of PC Magazine

ZDNet Products: DesktopUser