Intel Pentium 4 3.2E GHz "Prescott" Processor Review

Remember back in November of 2000 when the 1.5GHz Pentium 4 first came to light? Massive memory bandwidth, a juiced clock frequency, and countless architectural improvements promised a new era of performance. However, real-world tests on the fledgling processor revealed that in many cases, the 1.2GHz AMD Athlon, and indeed 1GHz Intel Pentium III, were still faster. Only later, as Intel improved the technology and augmented the chip's speed, did it become an attractive platform.

Today, the chip giant is in a similar bind. AMD has momentum following first a successful launch of its AMD64 initiative, and then the subsequent unveiling of its Athlon 64 3200+ and 3400+ processors. In order to keep up, Intel has had to drop the price of the flagship 3.2GHz Pentium 4, which is currently available for under $300USD ($512CDN).

Fortunately, Intel is in the position to begin manufacturing processors using a 90nm process, the very latest wafer technology that carries with it a host of other enhancements. Additionally, the Pentium 4 itself has gone under the knife, reemerging with larger caches, new instructions, and higher frequencies. The end result is a brand new core that, in theory, should be significantly faster than its predecessor.

Intel is also taking the necessary steps to ensure that the latest incarnation of its Pentium 4, formerly referred to as 'Prescott', has the capacity to scale much higher than the 'Northwood' core before it. Unfortunately, the initial effect on performance is significant enough to nullify the other architectural improvements.

Representatives from Intel believe that the passing of time will see Prescott assume the high-end position in its lineup, especially when software developers start optimizing for the processor's many new features. One of Prescott's sexist attributes is the technology on which it's manufactured. The new core represents Intel's first 90nm microprocessor, produced on 300mm wafers.

The smaller process technology provides Intel higher yields per wafer, increasing chip production and reducing costs. Whereas the Northwood core measures 146 square millimeters, Prescott is a petite 112. Consider that Northwood contains 55 million transistors to Prescott's 125 million, and you can see why manufacturing processes play an integral role in determining the complexity of a next-generation chip.

In the process of designing Prescott, the core moved from six layers of copper interconnect to seven, giving Intel more room to route electrical paths. Late in 2002, AMD used a similar tactic to give its struggling Thoroughbred core extra headroom, a move that took the Athlon XP from its 2200+ rating to 3200+.

Intel is also using strained silicon technology, which is a technique whereby silicon atoms are spaced further apart, resulting in less resistance as electrons flow through the channels of transistors. According to Intel, strained silicon consequently allows transistors to switch faster. Specifically, NMOS (negatively charged CMOS transistors) devices enjoy a 10 percent improvement in drive current, while PMOS (positively charged CMOS transistors) gain roughly 25 percent.

Keep in mind that Intel's roadmap calls for a 65nm process by 2005, which will also utilize strained silicon in addition to eight layers of copper interconnect using low-k dielectric material.

Pass the Cache

The massive jump in transistor count compared to Prescott's predecessor can be primarily attributed to the new core's larger caches. The Execution Trace Cache is still capable of storing 12k micro-ops; however, the L1 data cache has doubled in size from 8KB to 16KB. It is also 8-way set associative, as opposed to 4-way.

Prescott's L2 cache is two times larger than the Northwood core before it, weighing in at 1MB. It's attached to the processor core via a 256-bit bus, which, running at 3.2GHz, yields 102.4GB per second of bandwidth. As Prescott continues scaling, that number will follow suit. Incidentally, even while the Prescott is Intel's newest core, the Extreme Edition family, currently based on the server-oriented Gallatin core with 2MB of L3 cache, will persist as the top-end enthusiast product.