Intel's Prescott CPU--Behind the Scenes

Intel's Prescott CPU--Behind the Scenes

pentium5.jpgBy Gordon Mah Ung

In our March issue (“Why isn’t Prescott Pentium 5”, page 28), we take a deep look at Intel’s brand new CPU and find that, while feature-rich and full of promise for the future, the 3.2 GHz Prescott runs slower than Intel’s 3.4GHz P4 Extreme Edition.

In this exclusive feature for the website, senior editor Gordon Mah Ung explains how he ran his benchmark tests, and why the results didn’t surprise him.

For more info—including all of our findings and comprehensive explanations of Prescott’s new features and what they mean for end users, don’t miss our March issue--on sale now!

When Intel first briefed us on the Prescott, it was obvious that this new CPU would not initially be the performance king or Athlon 64 FX killer everyone expected. That’s because when Intel stacked up all three 3.4GHz CPUs, it still pushed the Pentium 4 Extreme Edition as the top dog.

Prescott and the Northwood Pentium 4 run pretty close in many tests, but there are some variations in applications due to the differences in the two CPUs’ cores. For our tests, Intel was only able to supply a Prescott clocked at 3GHz. Although the CPU was not clock-locked, internal governors that Intel put on its engineering sample chips prevented us from overclocking the CPU without overclocking the bus. Intel also couldn’t provide a 3.4GHz Northwood P4; but we already know that the stock P4 is shamed by the Pentium 4 Extreme Edition.

For our tests we ran a 3.2GHz Pentium 4 Prescott against the two fastest chips we had: a 3.4GHz Pentium 4 Extreme Edition and a 2.2GHz Athlon 64 FX-51. To gauge how Prescott performs in a clock-for-clock comparison, we also ran it against a 3.2GHz Pentium 4 Extreme Edition.

All our tests used the same ATI Radeon 9800 Pro card, Western Digital WD2500JB hard drives, Windows XP with SP1, Audigy 2ZS, and 1GB of Corsair Micro DDR400 RAM. Because the Athlon 64 FX-51 requires registered RAM, we used 1GB of registered Corsair Micro DDR400. All systems used the same Catalyst 3.10 ATI drivers and the same Creative Lab drivers. For the Pentium 4 system, we used an Intel-supplied D875PBZ motherboard and for the A64; an Asus SK8N mobo pulled the load. We originally started to use an Asus P4C800-E Deluxe board, but it exhibited quirks with the Prescott so we switched to the Intel Bonanza board. (Asus actually rates the P4C800-E as Prescott-ready, so we’re at a loss in trying to explain the problem.)

The big showdown was in BAPCo’s brand-new SYSmark 2004 benchmark. This benchmark loads the core of popular apps such as Word, Excel, Flash MX, Photoshop, After Effects, and WinZip to measure how fast a PC performs the task of creating and editing documents. While older versions have resided at the center of an acrimonious debate over the benchmarking suite’s fairness and balance, the new 2004 version has been created with the support of Intel and AMD (as well as Transmeta). SYSmark 2004 stresses the entire system: CPU, RAM, and hard drive.

In addition to SYSmark 2004, we also ran our standard set of application benchmarks: Premiere Pro, Photoshop 7.01, MusicMatch 8.1, and Mathematica 5. All four apps are gated by CPU performance. The predictable media encoding of MusicMatch 8.1 (and just about every MP3 encoder) favors Intel’s longer pipeline and higher clocks these days. While older versions of Premiere favor shorter-pipeline architectures such as the A64, PIII, and Centrino, Premiere Pro switched allegiances and is now a P4 lover. Finally, we ran Mathematica, which has traditionally favored CPUs brimming with floating-point performance; true to form, the Athlon 64 FX is the hands down winner here (and in SETI@Home).

To gauge synthetic memory bandwidth performance we used SiSoft Sandra 2004. This test used to be dominated by the P4 until AMD started putting memory controllers in its CPU cores. Now it’s Athlon’s world. Still, we wanted to see if the Prescott offered up any improvements.

To measure graphics performance, we used Specbench.org’s ViewPerf 7.1.1. ViewPerf tries to replicate what workstation-level applications like Pro/Engineer 2001, Lightscape, and DataExplorer display in OpenGL. While it’s more of an OpenGL test, with the same driver and videocard, we can see which CPU runs the view sets fastest. It’s also apparently quite an AGP bandwidth hog, so you can assess the strengths of a chipset’s AGP implementation.

Last, we ran a boatload of games and gaming benchmarks including 3DMark 2001 SE, 3DMark 03, AquaMark 3, Halo, Quake III Arena, Jedi Adademy, Comanche, and Unreal Tournament 2003. Other game benchmarks are available, but many of them are gated by the graphic card’s pixel shader abilities. Where available, we disabled audio and ran the tests at low resolutions to remove the graphics card from the equation.

0

Comments

+ Add a Comment

Log in to MaximumPC directly or log in using Facebook

Forgot your username or password?
Click here for help.

Login with Facebook
Log in using Facebook to share comments and articles easily with your Facebook feed.