Monday, June 12, 2000
A Performance Comparison of SMA Platforms.
VIA ProSavage PM133 vs Intel 810e.
Last year, Intel’s 810 and 810e chip sets established momentum for SMA (Shared Memory Architecture) in extremely low end PCs. This year the tide will begin to turn in the mainstream favoring higher performance and more flexible SMA platforms. These high volume mainstream platforms will require improved baseline graphics performance, PC133 support, AGP4x expansion, larger DRAM capacities, ATA66, AC97 and other I/O features.
With these requirements met, a single low cost Value PC platform can be used across the entire performance spectrum from low-end to high-end in the uni-processor desktop PC space. Of course, high end users will likely prefer an AGP upgrade, but if its 2D performance is adequate, an integrated graphics controller may satisfy large portions of the market.
From PC maker’s perspective, there is great advantage in the opportunity to upgrade or ‘upsell’ a competitive yet inexpensive base platform. A mainstream SMA platform will offer a low entry price with plenty of ‘build to order’ or retail upgrade headroom.
VIA’s ProSavage PM133 is the first chip set on the market to respond to these requirements. Essentially, the PM133 north bridge is VIA’s 694 north bridge chip integrated with an S3 Savage4 graphics controller. Intel’s offering this year will include the 810/810e plus the new 815 chip set. Essentially, the 815 is an 810e with PC133 support plus an external AGP bus. Though these two features are vital for any mainstream platform, the 815 will probably be hampered by the 810’s slightly under-performing integrated accelerator. VIA’s challenge will be to offer 815 or better performance while maintaining their reputation for lower cost.
Since Intel’s 815 is not yet released, this comparison will focus on the PM133 and Intel’s production 810e platform using PC100 SDRAM. In addition, to get an idea of the performance benefit that the 815 will gain from PC133 SDRAM, we will also do some background testing with the 810e overclocked to PC133.
VIA’s ProSavage PM133 reference platform was populated with two double-sided 128MB PC133 333 SDRAM DIMMs for a total configuration of 256MB. Similarly, an Intel CA810E production motherboard was populated with two double-sided 128MB PC100 222 SDRAM DIMMs for a total main memory configuration of 256MB. The CA810E board also includes a 4MB 32-bit 133MHz display cache memory subsystem for 3D graphics acceleration. In order to sample test the 810 chip set with PC133, we used a generic third party motherboard with the necessary overclocking capabilities.
We also had a strong interest to observe PM133 performance as a high-end platform (upgraded with an AGP accelerator). We upgraded the platform with a 32MB nVidia GeForce DDR AGP graphics card, and also switched to 256MB of PC133 222 SDRAM for main memory. Since the 810e platform is not upgradeable, we substituted Intel’s VC820 motherboard with 256MB of PC800 RDRAM. The GeForce accelerator could then be installed.
All systems were configured with 733/133 Pentium III Coppermine processors and 20GB Seagate Barracuda ATA66 hard drives. Windows98SE was installed and defragged on each system with the latest graphic, IDE and chip set drivers available as of the end of May 2000.
Business Applications Performance
To properly evaluate business productivity performance, we will look at three types of tests. Using each chip set’s integrated graphics controller, we will test Sysmark 2000 for media streaming and business software, Office Bench for MS Office application performance, and ZD’s WinBench for 2D business graphics performance
SysMark is an excellent starting point for overall business application performance analysis. This benchmark delivers consistent results across a broad spectrum of popular applications. Using 810e performance as a baseline, the ProSavage performance delta is charted individually for each application in the diagram below.
ProSavage clearly dominates the 810e in eight out of these twelve applications. The 810e wins two by about 1%, and in two more cases benchmark scores are essentially equal. Overall, ProSavage delivers a 2.3% performance advantage.
The three applications that show the greatest performance advantage on ProSavage are the media streaming applications - Naturally Speaking, Adobe Premier and Windows Media Encoder. Focusing on these media streaming applications, ProSavage delivers an average performance advantage of more than 6% (perhaps equal to or greater than one CPU speed grade).
Scores reported by Office Bench seem to corroborate and magnify some of the results reported by SysMark above. Office Bench exercises MS Office 2000 applications in a single tasking and multitasking manner.
ProSavage takes the honors in Office Bench with an overall performance margin of 7.6%. Even the Multitasking stress test favors ProSavage by a margin of nearly 10%. When comparing Sysmark and Office Bench results, the winners and losers for each application are the same, but the performance delta between platforms is about 5X greater in Office Bench than in SysMark 2000.
Under SysMark, MS Office apps deliver performance deltas ranging from –1.2% to +2.6%, a total spread of 3.8%. In Office Bench, these same applications range from –4.8% to +16.8%, a spread of 21.6%. We have observed that Office Bench is more dependent on 2D graphics performance than SysMark - thus the magnified performance differences shown here are largely attributable to the respective 2D accelerators.
WinBench 2000 – 2D Graphics
ZD’s WinBench is regarded as one of the best 2D business graphics benchmarks available. The graphics WinMark tests play back an intense sequence of GDI commands issued from various popular business applications. This approach stresses the 2D accelerator to the extreme in order to isolate graphics performance strengths and weaknesses.
Here again, ProSavage takes the lead by more than 10% and 18% respectively for business graphics and high-end graphics. This would seem attributable to the superior and more modern graphics accelerator engine integrated in ProSavage, along with its superior memory bandwidth.
Outlook for the 815
By overclocking the 810e to PC133 SDRAM, we attempted to simulate the performance of the upcoming 815 under business applications and 2D graphics loads. In most cases the performance change was very small, and even negative in some cases. Changing to PC133-333 from PC100-222 at times yielded a 2D graphics performance reduction. By moving to PC133-222, a small performance improvement could at times be identified, but not enough to make a significant dent in VIA’s performance lead with ProSavage. A more precise performance analysis must be withheld until after the release of the 815 chip set.
Game 3D Performance
Quake3 is without question the most popular game and gaming benchmark available in the industry. It is known to be fill rate limited in all but the lowest resolution modes. Quake3 is known to be a great 3D accelerator test, but not necessarily a good CPU performance test.
We tested the integrated graphics accelerators in NORMAL mode at 640x480x16. In addition, resolutions were increased to 800x600x16 and 1024x768x16 without changing any other image quality attributes from the otherwise NORMAL settings.
Under Quake3, VIA’s ProSavage delivers a remarkable 52% to 74% 3D performance advantage over Intel’s 810e. Importantly, performance at 640x480 and 800x600 both exceed 30fps, seen by many as a minimum threshold for playability. By contrast, the 810e does not achieve 30fps in any of these configurations. This raises serious doubt about its suitability for even casual gaming on a value PC.
By overclocking the 810e to 133MHz SDRAM (even using the fastest 222 SDRAM), frame rates increased by about 6-8%. This is still not enough to make a dent in VIA’s lead - even though the PM133 was configured with 133-333 SDRAM. When the 810e was configured with 133-333 SDRAM, its performance was basically the same as with PC100-222 (in the chart above).
In addition we experimented with turning off the 4MB display cache on the 810e motherboard. This had a negative performance hit of another 20-25% on the 810e motherboard. With both platforms configured equally, using low cost PC133-333 and no display cache, VIA’s PM133 delivers Quake3 frame rates that are generally 2x as fast as the 810e.
TreeMark 3D Geometry & Rendering
TreeMark is a simple benchmark provided by nVidia as a T&L and geometry stress test. In the absence of T&L acceleration, this benchmark presents a heavy CPU load in combination with a moderate 3D rendering load to the graphics accelerator and DRAM.
Again ProSavage takes a strong lead with a margin of about 25% over the 810e. Though these frame rates are not meant to be playable, they are an indication of how well each platform performs under a mixed load of geometry and 3D rasterization.
3D WinBench is a comprehensive performance tool conducting 53 standard performance tests. Unfortunately Intel’s 810e platform inexplicably and uniformly failed to run or produce a score in 41 of those tests. Though the 810e passed most of the image quality tests, it would not complete enough of the performance tests to permit a comprehensive 3D rendering performance analysis. We currently have no reason to believe that this is due to a configuration error, though we cannot help but wonder if the graphics driver is somehow conditioned to not run this benchmark.
The 12 tests that did complete are accounted for in the table below. Nine of the 12 are composited into a single Processor Test score. These tests use a graphics NULL driver in order to entirely eliminate any actual 3D rendering activity. The three WinMark scripts that did complete on the 810e were repeated at low and high resolutions in order to provide some kind of useful perspective on 3D performance.
Here again, ProSavage delivers a solid performance advantage that reinforces observations from all other 3D and 2D benchmarks used in this exercise. Under 3D WinBench, ProSavage exceeds 810e 3D performance by 33.1% at low resolutions, and by 18.5% at high resolutions.
Although the specific details are not shown here, in the 3D WinBench image quality tests, ProSavage passed 61 out of 69 tests (88.4%), while the 810e passed only 53 out of 69 tests (76.8%). So, in addition to offering substantially faster performance, one would expect ProSavage to also deliver a higher quality image.
3D Mark 2000
As a substitute for the numerous detailed 3D performance tests of 3D WinBench, we are including most of the tests from 3D Mark 2000 at the same two resolutions.
Without commenting on the details, it can be plainly observed that VIA’s ProSavage PM133 delivers an average performance advantage of about 25% regardless of resolution. This further supports the general observations made from Quake3 and 3D WinBench above.
PM133 as a High-End Platform
Though there should be many upgrade (or upsell) opportunities in a Value PC platform, the most obvious and perhaps most desirable is an AGP accelerator. Power users and enthusiasts will choose this path for better game performance and for improved business applications performance.
To satisfy this scenario, we have chosen to upgrade the platforms with the fastest AGP accelerator card commonly available in the market today – nVidia’s GeForce DDR.
Secondarily, performance conscious users have shown interest in DRAM performance upgrades as well. To satisfy this requirement, VIA’s ProSavage supports Virtual Channel SDRAM as well as PC133 222 SDRAM. For this benchmarking exercise, we have chosen the now widely available PC133 222 DIMMs. Since the 810e platform is not upgradeable in either of these two respects (AGP or faster DRAM), we are forced to compare ProSavage with Intel’s 820 platform instead. We used Intel’s VC820 motherboard populated with 256MB of PC800 RDRAM. All other platform elements were constant.
Using a few key benchmarks, our goal is to get an idea of whether or not ProSavage is competitive as a high-end uniprocessor desktop PC platform.
Game Performance: Quake 3
With the GeForce DDR accelerator installed, Quake3 frame rates skyrocketed in both platforms. Because we are testing for the high-end, we focused our effort on higher resolutions with 32-bit color and with other image quality attributes activated.
Even in these accelerator limited high-resolution modes, ProSavage sustains an attractive Quake3 performance advantage vs. the similarly configured VC820 platform. This performance advantage appears to expand slightly as screen resolutions increase, offering a greater performance advantage in the higher resolution modes most popular among gamers and enthusiasts.
Though the VIA performance advantage seems moderate, this 5-10% performance advantage is a pretty big deal to the power users who would purchase a high-end system configured in this manner.
Business Applications Performance: SysMark 2000
The SysMark performance of the two GeForce enhanced platforms are compared below to the baseline Value PC ProSavage (with SMA graphics only).
While the AGP upgraded platforms offer performance improvements in almost all cases, the overall performance boost is a surprisingly small 8% on average. Mainstream business desktop PC buyers may be perfectly content with an 8% performance compromise at a system cost savings of perhaps a couple hundred dollars.
It is particularly interesting to note that the ProSavage SMA configuration outperforms the VC820+GeForce configuration in Photoshop. This may be attributable to the efficiency of SMA 2D block transfer copy operations (memory to memory BitBLT rather than memory to screen BitBLT). This observation leads us to the question of what we might expect from the next generation of SMA platforms based on DDR SDRAM.
VIA’s ProSavage PM133 seems to have the ability to obsolete the 810 and 810e platforms. Severely lacking performance and upgrade options, the 810/810e may only hold on to its market share through aggressive price cuts and deal making on Intel’s part.
After a year of blood on the walls over Intel’s 820, RDRAM, the MTH and other chip set debacles, even previously staunch ‘Intel loyal’ PC buyers are softening to their alternatives. For many customers, the 810 was only taken seriously as an escape route from the 820 and as one of several options to cover Intel’s BX chip set shortages. As Intel continues to wean an unwilling customer base away from the legendary BX chip set, the market will seek a more flexible, feature rich, lower cost platform for the mainstream.
The acceptance of SMA into the mainstream offers the most significant system level cost reduction in recent history. Until other competitive chip sets arrive on the scene, the market will have to square off between Intel's 810/815 and VIA’s new ProSavage PM133. It will be interesting to watch.
By: Bert McComas, Inquest Inc.
Copyright © 2000 CST, Inc. All Rights Reserved