533MHz Monsters Come On Strong

The Lab begins a review of NT Alpha systems with a look at workstations from Aspen Systems and MaxVision

If you're used to working with Pentiums, Pentium Pros, and Pentium IIs that run at clock speeds from 75MHz to 333MHz, a 533MHz processor is a screaming machine. If you marry a 533MHz 21164PC Alpha processor to an Alpha PC164SX motherboard, you'll get a speed lover's dream workstation. Windows NT Explorer opens instantly on one of these workstations. If you add 128MB of RAM, Photoshop doesn't redraw multiple megabyte images one square at a time but pops them right onto the screen. And for 3-D engineering work, you'll be hard-pressed to find another system that's as fast for so little money.

To give you an idea of what NT Alpha machines can and can't do and what configurations vendors offer, I'll examine Alpha workstations in the Windows NT Magazine Lab over the next several months and report my findings to you. Most of the systems I will examine use the 533MHz processor and the Digital Semiconductor-manufactured PC164SX motherboard. The PC164SX motherboard uses a proprietary processor, the 21174, to interface the CPU, Level 2 cache, system memory, and PCI bus. The PC164SX has four PCI slots (two 32-bit and two 64-bit), two ISA slots with 1MB flash ROM on the ISA bus, two 9-pin serial ports, and support for 32MB to 1GB of RAM with 128-bit, 168-pin unbuffered Synchronous DRAM (SDRAM) Dual Inline Memory Modules (DIMMs). Most standard PCI and ISA cards are compatible with this motherboard's design.

64-Bit Processing Power
The 21164PC Alpha processor is available in 400MHz, 433MHz, and 533MHz versions. The 21164PC is a 64-bit all-RISC design (Pentium processor designs combine CISC and RISC). The 64-bit architecture means the 21164PC processes 64 bits of information at a time rather than the 32 bits that Pentium-class processors support. After the 64-bit architecture is widely adopted (Intel plans to release its 64-bit Merced chip in 1999), it will bring about a fundamental shift in technology, just as Pentium processors did when they supplanted the 16-bit x86 series. (See the sidebar "FX!32," page 80, to learn about running Intel-native NT code on an Alpha system.) With 64-bit systems, voice recognition will really work, artificial intelligence might not seem so artificial, and 3-D graphics and animation will be more impressive.

To feed its data-hungry silicon wafer, the 21164PC processor uses 8KB of Level 1 cache and 1MB of Level 2 cache. Level 1 cache is built into the processor and is usually very fast and rather small. You can find Level 2 cache on the processor or very close to it on the motherboard, as is the case with the 21164PC. Level 2 cache is larger than Level 1 cache and is not as fast, although it's considerably faster than the system memory. Both levels of cache store and arrange information for quick execution (using a method known as pipelining) when the CPU requests it. Pipelining supplies the CPU with a steady stream of data so that it does not sit idle.

A CPU accesses Level 1 cache first, Level 2 cache second, and the system memory last. Depending on the width of the internal (CPU) bus and the external (between the CPU and the motherboard) bus, a processor with Level 2 cache can transmit data transfers from memory to the CPU up to 10 times as fast as the same processor without Level 2 cache. This configuration is partly why companies didn't jump at the chance to integrate first-generation Pentium II processors into multiprocessor servers: Intel limited the Level 2 cache on early Pentium II processors to a built-in 512KB, which made these machines inferior in performance to Pentium Pros with a slower clock speed but 1MB of Level 2 cache.

Testing in the Windows NT Magazine Lab
To better understand NT Alpha workstations' overall system performance, I am coordinating the Alpha systems' Lab tests with AIM Technology (AIM) of Santa Clara, California. AIM is an independent company that promotes an open-systems approach to system-level Win32-compliant benchmarks through its load/mix modeling tests. AIM has provided benchmark services for the UNIX environment since 1981 and the NT environment since 1996. AIM uses 73 subsystem tests in a predefined application mix that models general workstation usage. AIM's tests cover all major subsystems in a multithreaded environment and provide two metrics for overall system performance, among other subsystem-level calculations.

The two most telling metrics that AIM's Workstation Benchmark produces are WNT Peak Performance and WNT Sustained Performance. The WNT Peak Performance test increases CPU, RAM, and disk caching to determine the maximum number of application jobs a system can process in 1 minute. The WNT Sustained Performance test measures the number of application jobs per minute at the point when a system is at the brink of failure. This value reflects the crossover point at which excessive application loads unacceptably hinder performance. I'll report the AIM WNT Peak Performance and WNT Sustained Performance metrics for every Alpha system I examine in this series.

Monitor resolution, bit depth, and the system's graphics card (2-D graphics cards generally give better AIM results than 3-D cards give) can affect the AIM test results. To level the playing field for the AIM tests on Alpha systems in this series, I will test each system at 800 * 600 * 16-bit resolution and with a Matrox Millennium II 4MB 2-D video card. I will install the Millennium II video card in each system specifically for the AIM tests if the Millennium II is not the system's standard video card. (For more information about AIM and to see the complete AIM benchmark test results for the Alpha systems I test, in addition to other Alpha systems, access the AIM Web site at http://www.aim.com.)

Every month in this series you can get a quick take on the test results for the systems by referring to the Benchmarks box, at the start of the article, which lists the two AIM parameters and the results of any additional tests. The test results for one system are most useful when you compare them to the results of other systems. Comparing system results lets you see how an Alpha system performs relative to other Alpha systems on the market.

Most of the Alpha machines I will review will have 1MB of Level 2 cache (however, I will test configurations with a Digital Semiconductor-designed motherboard that offers 2MB of Level 3 cache). To give you an idea of the kind of processing power an Alpha workstation is capable of, I ran Standard Performance Evaluation Corporation's (SPEC's) SPEC95 integer and floating-point benchmark tests on one of the review Alpha systems. SPEC is a consortium of vendors formed in 1988 to create metrics standards to differentiate systems through a series of benchmark tests.

When I ran the integer and floating-point SPEC95 CPU tests on the Lab's test 21164PC processor, I achieved results closely matching those published for the same processor on the SPEC Web site (http://www.specbench.org). The Alpha review processor achieved a SPEC95 integer value of 12.0 during my tests. (SPEC's benchmark integer value is 12.4.) The published SPEC95 integer value for a Pentium II 300MHz machine with the 440LX chipset is 11.7. The difference between the Pentium II and the Alpha unit I reviewed is insignificant, especially when you consider that the Alpha has almost twice the clock speed. However, the differences between the Pentium II and the review Alpha processor in SPEC floating-point values, which are important in calculations involving ray tracing, rendering, fluid dynamics, and finite element analysis, were significantly greater. The review Alpha processor achieved a SPEC95 floating point value of 15.4 during my tests. (SPEC's published floating-point value is 16.1.) The published SPEC95 floating-point value for a 300MHz Pentium II with the 440LX chipset is 8.15, only about half of the 21164PC's value. These SPEC95 test results indicate that if you perform multimedia work or serious number crunching, an NT Alpha machine will process your data more quickly than the fastest Pentium II.

Montrose
In its wide midtower box, Aspen Systems' Montrose Alpha 21164PC looks like any other workstation you'd pick off the shelf. However, with its clock speed of 533MHz, the Montrose is anything but typical. You get a lot of functionality for the price: a 533MHz Alpha CPU with 1MB of Level 2 cache, 32MB of SDRAM, two 64-bit and two 32-bit PCI slots, two ISA slots with 1MB of flash ROM, a Matrox Millennium II 4MB video card, a SCSI adapter, a Panasonic 8X SCSI CD-ROM drive, a Quantum 4.5GB Ultra SCSI hard disk, and NT Workstation installed. The Montrose workstation is a flexible, entry-level NT Alpha system--it's not intended to be a dedicated 3-D engineering workstation.

I tabulated AIM overall system benchmark results for the Montrose. The Montrose comes with the Matrox Millennium II 4MB video card (the Lab's Alpha-testing standard video card), and I set the Lab Alpha standard resolution of 800 * 600 * 16-bit on the Montrose's 14" monitor before I ran the AIM benchmarks. The Montrose scored 352.5 application jobs per minute on AIM's WNT Peak Performance load/mix modeling tests. The Montrose's WNT Sustained Performance value was 135.4 application jobs per minute.

If you compare the AIM benchmark values the Montrose achieved in the Lab to the benchmarks on AIM's Web site, the results look promising, especially when you compare the Montrose's lower price with the prices of other systems listed on the Web site. This comparison can be deceiving, however. I tested the Montrose workstation at 800 * 600 * 16-bit resolution, whereas AIM tested many of the systems on its Web site at 1024 * 768 * 8-bit and 1152 * 882 * 8-bit resolutions. When you compare the AIM benchmark results on identical systems tested at different screen resolutions, you're not exactly comparing apples to oranges. Rather, you're comparing red delicious apples to Granny Smiths: Yes, you will see a difference in test results, but it's a difference in degree, not in kind.

I have one little final point to make about the Montrose workstation (although sometimes the little things make a system attractive). I like the Axxion case on the Montrose workstation. It's wider than a standard case, and it comes apart in pieces. You can remove three screws on one side and pull gently on a handle that is formed into the side of the case to take the side of the unit off, without the struggle and hassle you often must go through when you open a unit. As I said, it's a little thing--but so is that 533MHz processor.

Symbion AXP164SX
I immediately liked MaxVision's Symbion AXP164SX workstation. The AXP164SX's thoroughbred 3-D graphics approach places this workstation a step above its 533MHz Alpha NT brethren. If computers were biological, nothing short of genetic engineering would have made this system as adept as it is at producing, managing, and moving 3-D images. The AXP164SX is one of a few workstations Parametric Technology has certified to run its high-end CAD/CAM design program, Pro/ENGINEER.

The Symbion system I tested came equipped to tackle 3-D graphics with 128MB SDRAM and two OpenGL-compatible 3-D video cards (you choose which one you want to buy with the system): the economical 3-DmaxP2, which uses 3Dlabs' PERMEDIA 2 graphics processor and has 8MB of Synchronous Graphics RAM (SGRAM), and the GLmax88M, with an 8MB VRAM frame buffer and an 8MB Enhanced Data Output (EDO) DRAM local buffer. The high-performance GLmax88M video card uses installable client driver architecture to optimize OpenGL performance with Pro/ENGINEER, SolidWorks, MicroStation, AutoCAD, 3D Studio MAX, Lightwave 3D, Softimage, and other rendering and visualization-intensive programs. Both video cards in the AXP164SX do a respectable job of running 3-D applications, but the high price tag and hardware-implemented Gourarud shading, texture mapping, depth buffering, anti-aliasing, and alpha blending on the GLmax88M set it apart from the lower-priced 3DmaxP2.

In addition to its high-end 3-D capabilities, the Symbion AXP164SX offers many amenities you want in a typical workstation. These amenities include two 32-bit and two 64-bit PCI slots; two ISA slots with 1MB of flash ROM; a 21174 core logic chip for interfacing the PCI bus and system memory; 1MB of Level 2 cache; and four 128-bit, 168-pin unbuffered SDRAM DIMMS slots capable of supporting 1GB of RAM. The Lab's test system included a 10/100 PCI Etherpower adapter, a Seagate 6.4GB SCSI hard disk, a SCSI adapter, multimedia sound card, Yamaha speakers, and a MaxVision 17" Max1769 monitor.

I wanted to test the Symbion's OpenGL-compliant GLmax88M video card, so I decided to use the Viewperf benchmark tests the OpenGL Performance Characterization (OPC) group released as an OpenGL benchmark. OPC is a nonprofit project providing comparison methods for OpenGL implementation performance across vendor platforms, operating systems, and windowing environments. (For more information about the Viewperf tests, and to compare the results I achieved with the Symbion AXP164SX with the results of other systems, visit the OPC Web site at http://www.specbench.org/gpc/opc.static.)

The most telling value in the Viewperf tests is the CDRS Viewset. The CDRS value in Viewperf measures a video card's modeling and rendering capabilities for CAD. The GLmax88M video card achieved a score of 26.546 on the Viewperf CDRS Viewset test. In another Viewperf test, the Data Explorer (DX) portion, which measures scien-tific data visualization capabilities, the GLmax88M video card tested at a 5.629 value. I also ran the Viewperf Lightscape test to measure the system's ability to accurately simulate global illumination effects. The GLmax88M video card tested at a value of 0.574 in the Viewperf Lightscape test.

I ran the AIM workstation benchmarks to see how the Symbion AXP164SX functioned as a system. I first tested the system with the GLmax88M OpenGL graphics card at 1024 * 768 * 16-bit resolution. The Symbion AXP164SX rated an anemic 161.1 application jobs per minute on AIM's WNT Peak Performance test. The system's WNT Sustained Performance benchmark was equally dismal at 110.4 application jobs per minute. Only when I installed the 2-D Matrox Millennium II 4MB video card in the Symbion AXP164SX and retested the system did the AIM re-
sults make sense. The new WNT Peak Performance value for the Symbion AXP164SX was 330.8 application jobs per minute, and the WNT Sustained Performance value was a respectable 191.8 application jobs per minute.

These tests make clear just how much a 3-D video card slows a system's performance. Unless you need extensive 3-D rendering and modeling capabilities, don't buy a high-end 3-D graphics card. The card not only won't do you any good, it will hamper your machine's performance on the applications you use most.