From today's techweb.com, 1 of 2
CPU Shoot-Out -- Intel is no longer undisputed champ
By Jonathan Blackwood
Intel may still be inside most PCs, but it just might get a little crowded in
there. Virtually the only game in town when it comes to microprocessors that
run Windows in native mode-that is, x86 processors-Intel has enjoyed a
playing field with little serious competition. But two old rivals-Advanced
Micro Devices (AMD) and Cyrix Corp.-are back with CPUs that could give
Intel a run for its money. u Sure, AMD and Cyrix were beaten in previous
attempts to dethrone Intel. But those efforts were stymied by fears about
compatibility and performance that lagged far behind that of the chips from
the Intel juggernaut. And no wonder. There's an old saw that says everyone
always talks about what a great dancer Fred Astaire was, but Ginger Rogers
took every step that he did-only backward and in high heels. The analogy to
the x86 market is apt: To avoid trespassing on Intel's intellectual property,
both AMD and Cyrix had to produce chips that behave precisely as Intel's
do, yet accomplish this feat with totally different chip designs. u To their
credit, both AMD and Cyrix have met the challenge. The compatibility issue
appears to have been solved, and new processors from both companies carry
Microsoft's designed for Microsoft Windows seal of approval. In our tests,
the AMD and Cyrix processors handled every program we threw at them
with aplomb. And the tests show the performance gap that helped keep Intel
comfortably ahead of this duo has narrowed considerably.
Lucky Seven
Socket 7 is the Zero Insertion Force (ZIF) connector that Pentium (and
Pentium-compatible) processors have plugged into on motherboards for the
past 18 months or so. Now computer manufacturers using Socket 7
motherboards have a choice of three processors using Intel's MMX
technology for their systems: AMD's K6, Cyrix's 6x86MX and Intel's P55C
Pentium. We compared the performance of these three 233MHz processors
in three different motherboards with three different chipsets, to find out which
is king of the Socket 7 hill.
Intel insists the future of the x86 platform is in its new Slot One architecture, a
patented connector that will hold only its new Pentium II processors. So we
also tested a 233MHz Pentium II system (a Dell Dimension XPS D233) to
see how its performance compared with the Socket 7 chips.
The Players
AMD introduced its K6 processor last April-the first 233MHz,
MMX-enabled processor on the market. Unlike the fifth-generation Pentium,
the K6 is a true sixth-generation processor (like Intel's Pentium Pro and
Pentium II designs). This means it supports dual pipelines, out-of-order
execution, branch prediction and other advanced technologies. It has dual
32KB on-chip caches, one each for data and instructions, for a total of 64KB
of level 1 cache. The K6 is also available in 166MHz and 200MHz versions,
with 266MHz and 300MHz chips expected by the end of the year.
Home for a K6 is a motherboard that supports 3.2-volt operation. Its other
requirements-support for a 3.5 clock multiplier and 66MHz bus-are relatively
ubiquitous these days.
Cyrix's 6x86MX is the newest entry in the Socket 7 sweepstakes. Until
recently, Cyrix was the only "fabless" provider of x86 chips, contracting with
IBM and SGS-Thomson Microelectronics to manufacture its chips. In late
July, National Semiconductor Corp. purchased Cyrix and will likely provide
Cyrix with manufacturing capability for the first time.
The 6x86MX is also a sixth-generation design, though it has a single unified
64KB on-chip cache. Cyrix uses a so-called "Performance Rating" (PR) for
the 6x86MX. The PR indicates the Pentium chip to which a Cyrix CPU is
roughly equivalent. The performance of the PR233 chip we tested is
equivalent to that of a 233MHz P55C, even though the 6x86MX actually
runs at 187.5MHz. It uses the same 2.8V required by Intel's P55C, and a
fairly common 2.5 clock multiplier; its only peculiar requirement is for 75MHz
bus operation.
Intel's P55C Pentium chip with MMX is the 800-pound gorilla in the Socket
7 marketplace. Because of Intel's dominant market position, it is the chip
against which the others must be compared. The P55C was the first
MMX-enabled processor on the market, but it is a fifth-generation design,
lacking features such as super pipelining, out-of-order completion and register
renaming. Its internal cache is the smallest of our group at 32KB (16KB data
and 16KB code). The P55C requires 2.8V operation, a 66MHz bus speed
and a 3.5 clock multiplier. Intel's sixth-generation processors, the Pentium
Pro (Socket 8) and Pentium II (Slot One), do not fit in Socket 7
motherboards. Because there is no 233MHz Pentium Pro currently available,
we did not include that CPU in our testing.
How We Tested
To ensure a consistent, controlled testing environment, we decided to test the
Socket 7 processors using three motherboards with three different core logic
chips, or chipsets. The three boards we chose were popular models: the
Baby AT form factor Hsingtech with VXPro chipset, provided by Way 2
Cheap Computers; First International Computer's (FIC) ATX form factor
PA-2011 with VIA 590VP chipset, provided by TigerDirect; and Asus
Computer International's ATX form factor TX97 Smart Main Board with
Intel's 430TX chipset, provided by Asus.
All three of the system boards shipped with 512KB pipeline-burst level 2
cache. We used 64MB of SDRAM, an STB Nitro 3D PCI video adapter
with 4MB of RAM, and a Samsung 2GB (nominal) IDE hard disk. We built
each system in turn, installed both Windows 95 and NT Workstation 4.0,
then installed our test suite: WINDOWS Magazine's Wintune 97, Microsoft's
Word 7.0 and Excel 7.0, Adobe Photoshop 4.0, Equilibrium's DeBabelizer
Pro and Autodesk's AutoCAD Release 14. We used the same hard disk with
each of the three boards. After running our tests on a given motherboard and
processor, we replaced the processor, made the appropriate jumper settings,
and repeated the drill.
All tests were performed at a video resolution of 800x600 pixels and a color
depth of 16 bits (65,000 colors). We ran each test three times (twice for the
lengthy multimedia test) and averaged the results.
With Wintune 97, we were most interested in a few key results: CPU
performance, measured in MIPS; cached-disk throughput, which measures
CPU performance as much as, or more than, disk quality or performance;
and video throughput.
Wintune's measurements are raw-or synthetic-benchmarks. This means the
program measures specific aspects of system performance, which are not
necessarily indicative of system performance on a typical application. For
example, when we tested the Dell Dimension XPS D233, a 233MHz Pentium
II with Intel's Accelerated Graphics Port (AGP), we noticed a big boost in
video throughput when using the standard STB Velocity AGP card as
opposed to the PCI-based STB Nitro 3D: 107Mpixels per second vs.
41Mpixels per second. Application performance in the two configurations,
however, was almost identical, with the AGP-based system registering slightly
(2 to 3 percent) faster results.
We were much more interested in how the systems-and the
processors-performed with actual applications. The suite of application tests
we use is designed to mimic the demands typical users make on their systems.
Our Word and Excel macros use various system functions, and make
extensive use of the video and hard-disk subsystems. Our multimedia
benchmark, which uses a script written in Equilibrium's DeBabelizer Pro to
execute a series of Adobe Photoshop 4.0 filters on a set of photo images,
exercises the processor more than the other subsystems. And our AutoCAD
R14 test, which opens and renders the program's CHEVY.DWG sample
drawing, is almost a pure floating-point application, though the compiling
method used by Autodesk favors the fifth-generation P55C from Intel.
We used 64MB of RAM for our tests with both operating systems, because
AutoCAD R14 is so memory-intensive that 32MB slows it to a crawl. (We
recommend a minimun of 32MB for Windows 95 and 64MB for NT.) Keep
in mind that most Socket 7 chipsets-including Intel's-only support caching of
the first 64MB of RAM. So installing more than 64MB will actually slow
overall system performance.
CPU Comparison
The specific, raw results of our tests are shown in the sidebar "Great
Performances." We also show the normalized results (see sidebar "System
Shoot-Out Scorecards"), in which each result is expressed as a percentage of
the best score in that category. We then derived overall normalized scores for
raw benchmarks, application performance and combined scores. There are
three categories in the synthetic benchmarks and four for application
performance. Combining these two sets of tests, a perfect total score would
be 7.0, assuming the top score of 100 percent in each category equals 1.0.
When we tallied up the scores, we found that each of these chips excelled in a
specific area.
The K6 was second only to the PII in the raw CPU benchmark, and it had
quite respectable application scores-better than the Intel P55C and slightly
worse than the Cyrix and PII on ordinary business apps. It almost equaled the
P55C on our multimedia test, but rendered the AutoCAD file about 10
percent to 20 percent slower than the P55C and substantially slower than the
PII. On the NT tests, which we did not run on the Pentium II due to the lack
of drivers for its AGP video card, the K6 was marginally the fastest
performer among the Socket 7 processors.
Cyrix, running at a lower clock speed, gives up some raw computational
power. Its best scores were in cached-disk performance, where its 64KB
unified cache can make a difference, and in ordinary business applications,
where its superior branch prediction and out-of-order execution schemes can
make up for what it lacks in raw processing power. Although the Pentium II
was faster on most measurements, the Cyrix chip performed better on Excel
performance and PCI video throughput, and was nearly its equal on our
Word macro (as was the K6) and cached-disk performance. Both Intel chips
stood out on the multimedia and AutoCAD render tests. The 6x86MX was
relatively slow at these two tests.
AMD likes to position its K6 processor as a direct Pentium II competitor,
but at 233MHz, its performance on most tests fell squarely between the
P55C and the PII. Still, it offered well-rounded performance at a variety of
tasks and with its upcoming 266MHz and 300MHz versions, it will indeed
provide a cheaper Socket 7 alternative to Intel's high-priced, Slot One
Pentium II powerhouses.
Cyrix's 6x86MX was an outstanding performer at ordinary business
applications. It also performed acceptably on multimedia and computational
tasks such as high-end graphics, but its performance in these two areas was
the worst of what is admittedly a lightning fast lot.
Intel's P55C, on the other hand, is the slowest of the Socket 7 chips on
typical business tasks, but the best of the bunch on multimedia and high-end
graphics. The PII excels in every area, though its business-application
performance was basically matched by the 6x86MX.
Intel's highly touted AGP made little difference in application performance in
our tests. Its strength is its ability to provide background textures for 3D
graphics, which currently is more important in games than in business apps. It
is, at present, a point of differentiation between Intel's Slot One architecture
and the existing Socket 7 motherboards. But AMD has announced its
intention to provide AGP support in new chipsets for Socket 7
motherboards, so AGP should be available on both platforms by year's end.
The Winner Is ...
The overall performance of the three Socket 7 processors is so evenly
matched that you wouldn't notice a performance difference among them.
Though each has its particular strengths, none stands out so dramatically from
the others. The PII is faster in most areas, but not fast enough to justify paying
a major premium to attain the extra performance, especially for business
apps.
AMD aims to price the K6 about 25 percent lower than the equivalent
competing Intel processor, although it's unclear whether the company thinks
the Pentium II or the P55C is its competition. Since it's unlikely that Intel will
tweak the P55C beyond 233MHz, AMD's pricing formula will undoubtedly
refer to the PII. Cyrix's strategy is to price the 6x86MX at half the cost of the
equivalent Intel chip. And Intel, always a worthy competitor, has recently
demonstrated its willingness to cut prices to meet the competition head on. As
of this writing, the current prices in unit quantities of 1,000 chips are $290 for
the 233MHz K6, $290 for the 233MHz 6x86MX, $337 for the 233MHz
P55C and $455 for the 233MHz Pentium II. That translates roughly to
system prices of $2,200 for the K6, $1,500 for the 6x86MX, $2,500 for the
P55C and under $3,000 for the Pentium II.
It will likely take some time for the AMD and Cyrix chips to find their way
onto corporate desktops, though Digital's new K6-based Venturis FX-2 on
our WinList is one system intended for the corporate market. But
less-conservative buyers now have a range of chips to choose from, at
better-than-ever prices. The slugfest among these three chip giants may leave
them battered and bruised, but for the consumer the good times are just
beginning.
SIDEBAR: Great Performances
We found no significant differences in performance between the three
motherboards and chipsets we used for our test beds. Because the chips
performed similarly on all three boards, we combined the average scores for
each test bed to compare performance. We included the 233MHz Pentium
II-powered Dell Dimension XPS D233 for comparison purposes, though it
uses a completely different motherboard design (Slot One) and offered
Accelerated Graphics Port (AGP) graphics via an STB Velocity 3D AGP
video adapter. We also tested the Pentium II system with the STB Nitro 3D
PCI card used in the other tests.
All processors were tested with Windows 95 and Windows NT Workstation
4.0. We did not perform Windows NT tests on the Pentium II system,
because no video drivers are yet available for the STB AGP card.
SIDEBAR: Add Muscle to Old Systems
Socket 7 isn't the exclusive domain of the high-powered 233MHz chips from
Intel, AMD and Cyrix. It's been around for a while and has been in use since
the introduction of the 0.34-micron, 75MHz P54C Pentium two years ago.
So you may be able to upgrade an older machine-all the way up to a
233MHz AMD-K6.
But you can't just drop any K6 into your old 75MHz Pentium motherboard.
For one thing, it's doubtful many of those older systems will support the
3.2-volt operation the K6 requires.
Evergreen Technologies may have a solution. The company's $499 MxPro
Processor for Pentiums offers a viable upgrade route. Evergreen has provided
CPU upgrades for some years now, and the experience shows: This product
comes with a (flash) BIOS upgrade, ample documentation, a voltage
regulator and a heat sink/fan assembly.
We upgraded a two-year-old Gateway 2000 P5-75 system. Evergreen has a
long list of systems that work with its product, with more added weekly. You
can check the list at a computer retailer or on Evergreen's Web site
(http://www.evertech.com). If everything works right the first time, you can be
up and running in less than an hour.
Our test system had 32MB of RAM, a 720MB Western Digital Caviar hard
disk and a Matrox Millennium video adapter. The results of the upgrade are
shown in the "Before and After" table.
The results were stunning. As a rule, one system should be twice as fast as
another before you'll notice a significant difference in performance. This
upgrade certainly meets that criterion (a 200 percent improvement means that
the upgraded system is three times faster than the original). The upgraded
system comes within striking distance of a new 233MHz system.
Evergreen's MxPro Processor for Pentiums can be purchased at retail outlets
and from many direct merchandisers. Upgrades to 200MHz ($349) and
166MHz ($199) are also available. Contact Evergreen Technologies at
541-757-0934 or at its Web site. |
