First Tests: Intel's 3-GHz Pentium 4
Date 14-Nov-2002 13:24:32
|Hyperthreaded desktop CPU delivers mixed results, but systems embrace it.|
Intel is launching its 3.06-GHz Pentium 4 with hyperthreading technology, and major PC makers are readying PCs powered by the new CPU. In coming days and weeks, Dell, Gateway, and Hewlett-Packard, as well as specialty vendors Alienware, Falcon Northwest, and Xi, will roll out high-end products featuring the chip.
Intel says hyperthreading lets the P4 act as two processors, simultaneously tackling multiple applications or a single application with multiple threads. That should boost performance, but exclusive PC World tests show mixed results at best.
Depends on the Task
Tests on three of the first systems to use the new chip indicate that hyperthreading's benefits are largely application- or even task-specific. For example, on the office applications most people use, such as Microsoft's Word and Excel, hyperthreading's impact was neutral or slightly negative.
Even on specialized tests designed to highlight the benefits of hyperthreading, most improvements are negligible. Exceptions are select filters in Photoshop, some aspects of video and photo programs like Premiere and VideoWave, and certain multitasking tests.
PC World's results also show that a PC equipped with AMD's Athlon XP 2800+ CPU, which runs at 2.25 GHz, more than held its own against these 3-GHz systems running 800 MHz faster. Hyperthreading is still new, however, and Intel says you should see improvements as developers modify their applications to take better advantage of the technology.
PC World tested three shipping 3.06-GHz P4 systems, each containing 1GB of PC1066 RDRAM and running Windows XP Pro. (For now, hyperthreading works only with Windows XP and Linux 2.4.x, which can recognize the technology and use it as they would physical dual processors). The P4 systems tested included Falcon Northwest's Mach V ($3288), Gateway's 700XL ($4158), and Dell's 8250 ($3419). Each PC was loaded with high-end components and ready for the most demanding computing tasks.
The shipping comparison machines consisted of Falcon's Athlon XP 2800+ Mach V with 1GB of DDR333 SDRAM ($2786) and Xi's MTower DPR with two Xeon CPUs and 512MB of PC800 RDRAM ($3924). Both comparison systems also run Windows XP Pro; all Intel systems were tested with hyperthreading both on and off.
On the PC WorldBench 4 test suite, hyperthreading showed little effect. That's no surprise, given that the multiprocessor-aware OS imposes its own overhead, and Intel predicted that standard business applications (such as those in PC WorldBench 4) would show little performance gain when run individually.
The Athlon XP system scored 130 on PC WorldBench 4, besting all three P4 systems as well as the dual-Xeon unit. In fact, the score for each P4 system was either the same or slightly worse with hyperthreading turned on; the Dell experienced the biggest drop, scoring 117 with the feature disabled and 113 with it enabled.
More troubling, two of the three 3.06-GHz PCs scored 7 to 8 percent below the average of 11 older systems tested. Five of those older P4 systems, running at 2.53 GHz, averaged 121 on PC WorldBench 4. Six 2.8-GHz P4 systems averaged 123.
To evaluate hyperthreading technology more thoroughly, the PC World Test Center devised several new tests. Analysts created two Photoshop tests, engineered two multitasking tests, and used snippets of the still-in-progress PC WorldBench 5 (due out in 2003).
The first Photoshop test involved 20 common filters, most of them specified by Adobe or by PC World's own art department, and the others chosen from an array of Intel recommendations. Because Photoshop is a multithreaded application, some improvements might be expected from using Intel's new technology with it. But on our first series of Photoshop tests, the Athlon XP PC beat the P4 and Xeon systems, regardless of hyperthreading status. The Dell PC saw a 2 percent boost from hyperthreading--the largest of the P4s here, but hardly noticeable to most users.
A second round of Photoshop tests used 20 filters nominated by Intel specifically to showcase its technology. Running these filters, which incorporate numerous intermediate calculations that hyperthreading can run in parallel, the P4 systems showed marked improvement. For example, the Dell took 118 seconds to complete the test with hyperthreading turned off (slower than the Athlon XP unit's 104 seconds), but its time dropped to 95 seconds with hyperthreading on--a significant 24 percent improvement.
Tests of VideoWave used both DivX-based files and MPEG-1 files. On the DivX files, scores fell slightly or remained the same when PCs had hyperthreading enabled; but on the MPEG-1 files, they improved by up to 12 percent.
Hand-timed multitasking tests yielded a mixed bag of results for the P4 systems. One test involved having the photo-management application ACDSee convert 315 TIF files into JPEGs in the foreground while McAfee Viruscan checked a directory of folders in the background. Though all three P4-based systems performed these tests better than the Athlon XP PC did, the benefits of hyperthreading were far from uniform. For example, the Dell experienced a 4 percent slowdown on ACDSee but a 4 percent boost on McAfee, whereas Gateway netted an 8 percent McAfee boost and no slowdown on ACDSee. In each case, however, running the applications consecutively took less time than running them concurrently.
The opposite results occurred in a second multitasking test. Here, Ahead's Nero Burning ROM re-created a 490MB hard drive image in the foreground while MusicMatch converted ten WAV files into MP3s in the background. In this case, running the two applications simultaneously went faster than running them separately--by about 30 seconds for the P4s and by 53 seconds for the dual-Xeon PC. Again, the P4-based systems performed best overall, and most experienced a slight performance boost with hyperthreading in the concurrent runs.
Bottom line: Hyperthreading's effects vary wildly, and depend greatly on what application--or set of applications--you run. In most cases, however, its effect was minor. And it would be difficult to predict how your favorite applications will react, on the basis of these tests, due to the variability of the results.
Why would a faster chip capable of acting like two processors prompt such small change in test scores?
"One ill-behaved single-thread application could cause problems in a hyperthreading PC," says Kevin Krewell, general manager at research firm MicroDesign Resources. While the new P4 tricks the operating system into believing it is two chips, it still has the resources of only one.
For example, the new P4 shares on-board cache as needed, but it splits its write buffer resources in half--so a single-threaded application that once had access to all eight write buffers can now only use four, he explains.
Brian Fravel, Intel's desktop marketing manager, acknowledges that, as PC World's tests demonstrated, some single-threaded applications experience no performance gains and may even incur losses under hyperthreading.
"In some cases it's because it's poorly written code, and in some cases it may be an app that is very focused on a single unit--integer or floating point," Fravel says. But few apps should run noticeably slower, he adds.
The key to taking advantage of hyperthreading is not simply to run multiple programs or a multithreaded application, but to make sure that those programs hammer the CPU with tasks, says Dean McCarron, principal analyst at Mercury Research. "Both threads need to be throwing instructions at the processor simultaneously," he says.
Software doesn't need to be optimized for hyperthreading, but vendors can tweak their programs to take advantage of it, according to Intel's Fravel. Intel won't disclose specifics, but Fravel says that makers of image and video editing applications are likely to be among the first to optimize their programs.
"When you look at video encoding and transcoding today, it's so processor-intensive--there are so many places that if you write it to use hyperthreading you can take advantage of those unused processor cycles," Fravel says.
Video game programmers could make hay with multithreading, too, building more sophisticated artificial intelligence engines while simultaneously creating a more realistic 3D environment, he adds.
Intel clearly hopes hyperthreading will push PC sales just in time for holiday promotions, says MDR's Krewell.
"If you look at the timing, Intel is clearly trying to goose the market," he says. "Intel will have the latest and greatest chip with the newest buzzword and that could be a substantial marketing tool."
Don't expect the competition to sit still as Intel trumpets its new technology, however. AMD has taken a beating from Intel on processor speed recently, but it has new ventures in the works.
AMD plans to launch an improved version of its current Athlon XP chip, code-named Barton, as well as its 64-bit Hammer-based processors, in the first half of 2003. According to Krewell, the highly efficient Hammer, which includes onboard memory controllers, should compete vigorously against Intel's hyperthreaded P4.
In the meantime, for straight-ahead performance and price, you may be better off choosing one of Intel's older P4s or a high-end Athlon XP-based system. If you are serious about multithreaded applications, heavily use video or photo editing programs, or multitask as a matter of course, the new P4 systems may suit your needs, if you choose applications carefully.