Intel Turns Up the Heat on AMD

PALM SPRINGS, CA - Palm Springs is east of Los Angeles, right in California desert country. So Palm Springs in September vaguely resembles an oven. It's an apt metaphor as Intel released information on its upcoming Katmai CPU, hoping to turn up enough heat to wilt competitor AMD.

Katmai is Intel's successor to the Pentium II. Slated for release in early 1999, Katmai's main improvements center around two areas - improvements to the internal instruction set for faster 3D graphics and technology to move that 3D data to and from memory much faster than before.

MMX on SteroidsMMX brought a new technology to computers called SIMD. SIMD stands for single instruction, multiple data. In essence, MMX allowed one CPU instruction to act on up to eight data items at a time, instead of just one item at a time. This was particularly useful for matrix math - the kind of arithmetic needed to handle MPEG data, audio data, and some 2D graphics.

However, MMX had some serious flaws. For one thing, it could operate only on integer data - whole numbers. This precluded any improvement to 3D graphics performance by the CPU. MMX also shared the floating-point registers with the floating-point unit. (Registers are temporary storage locations in the CPU itself. CPU instructions operate on data in the registers.) There was a switching delay when changing from MMX data to standard floating-point information. The Pentium II improved on this somewhat, but it's still there.

Katmai New Instructions fixes these flaws. First, there are new 128-bit-wide registers. They allow up to four 32-bit single-precision floating-point data items to be operated on simultaneously by one instruction. Intel calls this SIMD-FP (single instruction, multiple data floating-point). This is just the kind of data type used in the math used to calculate geometry and lighting for 3D games. Secondly, these registers are independent of the classic floating-point registers and allow easy intermixing of SIMD-FP data and more standard IEEE-format floating-point items. This can lead to significant improvements in geometry calculations. In his presentation on the 3D technology used in Half-Life, Greg Newell of Valve suggested that a "70-100 percent improvement is conservative."

Michael Hara of nVidia agreed. He noted that artists put in a lot of work on today's 3D games, particularly on the texture side, to bring detail to 3D game objects that really have very little detail. Katmai, coupled with third- and fourth-generation 3D graphic accelerators, will allow for much more real detail to be there - more polygons and smaller polygons, as well as a richer texture set. The net result is that games will look better and play better. As Hara put it, "Today's 3D graphics will look like yesterday's 2D graphics compared with what will be out by Christmas 1999."

A Fire Hose for DataBeing able to generate all those cool 3D graphics is very nice, but you have to get information to and from memory. The kind of data used by 3D graphics is often bigger than even the cache in a Xeon processor can hold, so there is lots of data being moved in and out of main memory. Katmai improves this transfer to and from memory by incorporating technology Intel is calling streaming memory.

Here's the problem: CPUs are getting faster and faster. Memory isn't keeping up. Even 125MHz SDRAM is being eclipsed by the future of 500MHz and faster CPUs. Intel has placed its blessing on a new memory type called Direct RDRAM (Direct RAMBUS DRAM). RDRAM improves the transfers of memory to and from the CPU up to 300 percent. Of course, support for RDRAM will require new motherboards and chipsets, but the performance improvement will be even more substantial than the shift from EDO to SDRAM.

The Bottom LineFor those users gloomily contemplating their spanking-new 450MHz Pentium IIs, the good news is that the first Katmai CPUs will run on 440BX-based motherboards. And the streaming memory technology will even improve the movement of data in and out of main memory for SDRAM, though not to the extent of RDRAM.

Also, it may be some time before games really take advantage of Katmai's new instructions. It's akin to the problem facing K6-2 users who have been waiting for 3DNow support. Until games support the transform and lighting engine in DirectX 6.0, there won't be much improvement in game performance (other than the obvious clock-rate improvement). OpenGL drivers will have to be revved as well. So users rushing out to buy the first Katmai systems may only see modest improvements. Rather, Katmai is a hedge to the future. When those Katmai-enabled games come out later in 1999, 3D games will look richer, more detailed, and run better than ever before.