More than even AMD, Apple Computer loves to brag about and sling benchmarks around to tout "Pentium-crushing performance." But as its desktops' PowerPC G4 processors inched up to and finally past the 1GHz clock-speed barrier while Intel's CPUs tripled that, both perception and reality were rapidly moving Windows PCs' way.

Two weeks ago, Apple made a bold move to recapture multimedia and publishing professionals' attention: introducing new Power Mac G5 systems featuring what it calls "the world's first 64-bit desktop processor," the "revolutionary" PowerPC G5 -- with 40 percent higher clock speed, double the on-chip cache, and over six times the bus bandwidth of its fastest G4 predecessor.

Besides moving from 32 to 64 bits, the G5 represents Apple's turning from longtime supplier Motorola to its other partner in the PowerPC trio, IBM -- the new CPU is actually the PowerPC 970 that IBM unveiled in October 2002. Here's a brief look at the processor and how it draws on Big Blue's flagship server technology to improve on previous PowerPC desktop engines.

Downsized from the Enterprise

Built using IBM's 0.13-micron copper and silicon-on-insulator (SOI) process at its 300mm wafer fabrication plant in East Fishkill, N.Y., the PowerPC 970 is based on the IBM Power4 server CPU -- a 64-bit juggernaut that incorporates two CPU cores in one chip, along with a shared 1.5MB Level 2 cache and Level 3 cache controller.

To tame this beast for desktop duty, IBM removed one of the processor cores and the L3 controller and trimmed the Level 2 cache to 512K, accompanied by 96K of Level 1 cache (32K for data and 64K for instructions). It also replaced the Power4's exotic chip-to-chip connection circuitry, which seamlessly supports the joining of four Power4 chips into one eight-core, 680-million-transistor, 500-watt multiprocessing module, with a more conventional (but still formidable; more on that in a moment) bus interface unit.

The resulting CPU, introduced at clock speeds of 1.6GHz and 1.8GHz and recently stepped up to 2.0GHz (Apple's Power Mac G5 line offers single-CPU configurations with the first two chips and a dual-2.0GHz flagship), squeezes 58 million transistors into a 118-square-millimeter die -- fractionally smaller than Intel's "Northwood" Pentium 4 -- packaged in a 576-pin ceramic array.

Like its competitors, it's a superscalar processor -- able to execute five instructions per clock cycle (four plus a branch instruction). That's one up on Motorola's most recent 32-bit G4 core, but where the G5 really pummels its predecessor is in pipeline length: Instead of the original G4's 4-stage or G4+ chip's 7-stage integer pipeline, the PowerPC 970 follows the superpipelined approach of the 20-stage Pentium 4 with a 16-stage integer pipeline -- 21 stages for floating-point instructions, as many as 25 stages for single-instruction-multiple-data (SIMD) multimedia instructions.

The 970 has 10 function units -- two integer, two floating-point, two SIMD, two load/store, a branch unit, and a condition-register unit. The combination of breadth and length yields what Apple likes to call "massively parallel computation": as many as 215 "in-flight" instructions being juggled at any given instant, compared to 126 simultaneous instructions underway for the Pentium 4 and a mere 16 for the original PowerPC G4.

Bandwidth Bonanza

While the 32-bit G4 had been tweaked with a 36-bit memory addressing mode (giving a theoretical 64GB rather than 4GB memory ceiling), the 970 can handle more memory than any Power Mac G5 owners are likely to install. The processor matches its 64-bit data path width with a 64-bit virtual memory-address range -- supporting a mind-blowing 18 exabytes or 18 million terabytes -- along with a more than ample 42-bit or 4TB real address range. (By contrast, AMD's Opteron has a 48-bit virtual and 40-bit physical address space.)

The interface between CPU and memory is perhaps the PowerPC 970's biggest advance from its predecessor. While AMD's Athlon XP 3200+ boasts a 400MHz and Intel's latest Pentium 4 processors an 800MHz front-side bus, most G4 processors -- including those in current Apple iMac and eMac models -- hobble along with a 133MHz bus, and even Apple's fastest dual-G4 towers peaked at a front-side bus speed of 167MHz yielding peak bandwidth or data throughput of 1.3GB/sec (about one-fifth that of the current Pentium 4).

IBM's design for the 970 blows away that bottleneck with a system bus -- technically, two 32-bit unidirectional buses or point-to-point interfaces -- running at an effective clock rate of half the CPU clock, such as 900MHz for the 1.8GHz processor or 1GHz for the 2GHz chip. The latter yields 8GB/sec of raw aggregate bandwidth, or a still formidable 7.1GB/sec after setting aside some capacity for address and control signals.

Finally, the PowerPC 970 resembles AMD's Opteron and Athlon 64 in being a 64-bit CPU that offers seamless compatibility with existing 32-bit software (of course, PowerPC rather than x86 software). In addition to the native 64-bit mode, with the 42-bit real and 64-bit effective addresses mentioned earlier, the processor supports a native 32-bit mode, which clears the top half of all effective addresses and uses the first 16 entries of the segment look-aside buffer as on-chip segment registers, reducing the number of off-chip table accesses.

The 970's support for SIMD instructions represents another change from its Power4 parent -- IBM added the AltiVec (which Apple calls Velocity Engine) vector-processing extensions added to the PowerPC instruction set by Motorola. They're akin to Intel's SSE2 or AMD's 3DNow Professional variations on the x86 theme, though optimized for the graphics programs such as Adobe Photoshop popular with Apple's art-director audience. (See CPU Planet's "Multimedia Extensions" article -- Ed.)