November 6, 1997

EDN HANDS-ON EVALUATION

HAS WINTEL WON THE WORKSTATION WAR?

MAURY WRIGHT, TECHNICAL EDITOR

More and more workstation vendors are moving to Intel µPs and Microsoft OSs, thanks to the cost advantage of the high-volume PC business. This Hands-On report reviews a few Wintel workstations, both from workstation-industry stalwarts and from PC vendors moving into the high-end business, and helps you tell what differentiates a PC from a workstation.

Last year, I predicted that Intel-based Windows NT systems would encroach on the turf of traditional workstation vendors, such as Sun (Palo Alto, CA) and Hewlett-Packard (Reference 1). Although many workstation vendors and analysts considered this conclusion implausible, I was right. Despite their proclamations that Intel-based NT systems would succeed mainly as upgrades for users that could previously afford only PCs, a current unmistakable trend toward Intel (Santa Clara, CA) and Microsoft (Redmond, WA) Windows NT is taking place.

For example, industry stalwart Hewlett-Packard is moving toward an Intel-based hardware platform that will support a choice of NT or Unix. Moreover, at the annual Design Automation Conference in June, HP revealed that most of its EDA-software partners requested NT demo machines to display their wares. The company has also announced its Kayak family of NT workstations and merged its PC and workstation organization.

The trend has the momentum to lead the so-called Wintel architecture to a dominant position in the EDA market. With significant growth evident in that market, EDN now takes a hands-on look at the emerging class of workstations. Although not exhaustive, this investigation will help you understand the performance potential of these machines running sample EDA software, get a glimpse of the marvelous mechanical design of some representative systems, and learn what distinguishes workstations from PCs (see box "PCs vs workstations").

To help evaluate workstations, I spent some time with representative systems from Compaq, HP, and Intergraph. I talked with several vendors, including Dell and Intergraph, about new systems that aren't yet shipping. I also spent some time with a Digital Alpha-based Windows NT workstation and found that Alpha systems may ultimately prove to be the only obstacle to Intel's domination of the workstation market.

You might ask what transpired in a year to make Wintel and EDA a sure bet. As in many markets, including embedded systems, the volume of the PC market has created a significant value proposition for other types of applications that can leverage similar technologies, including µPs, bus interfaces, data-storage technologies, and graphics. It's not just price that makes Wintel a winner, however. The new Pentium II µP matches or exceeds the performance of any mainstream µP, except the Alpha. When you consider the price advantage of the Pentium II, you will have a tough time making a case for other workstation architectures.

Unix: still kicking

Now, don't misunderstand the situation: Unix will be around for a long time, if not forever. For the foreseeable future, Unix will dominate at the high end of the workstation business. Typically, application servers that host shared resources, such as high-end simulators and synthesizers, will stick with Unix. Windows NT still lacks security features and the stability to run months without a reboot--characteristics that mission-critical application servers require.

The engineers' window to these servers and applications, however, will more often be a Wintel workstation. Moreover, workstation EDA-software vendors, such as Cadence (San Jose, CA), are rushing to port their interactive tools, such as schematic-capture packages, to the Wintel architecture. Meanwhile, PC EDA-software vendors, such as Accel Technologies, are rewriting their Windows packages into full-fledged 32-bit applications that run equally well under Windows 95 or NT.

With a number of vendors offering Wintel workstations, you will face a tough task choosing the one that is best for your application. You typically want to buy the best performing machine that your budget allows. Depending on your application, however, you might be better off with a dual-processor system rather than a traditional single-processor system. In other cases, a single-processor system with high-speed memory, such as synchronous DRAM, may prove a better option. You should also consider mechanical design before making a decision.

Although the systems I tested are wicked-fast, mechanical features made a lasting impression on me. I regularly experiment with new add-in cards and peripherals and have been guilty of leaving the cover off a system for years on end. The Compaq Professional Workstation 6000 may be the first system I've used that I wouldn't mind leaving the cover on. The tower system includes two card cages that you can remove from the rear in a snap. One of the cages can accommodate two Pentium II processors that Compaq supplies on small pc boards along with the necessary dedicated power converter. The card cage also includes 12 DIMM sockets for memory expansion. You can perform a memory or processor upgrade easily in less than a minute without tools.

The second card cage includes six PCI slots, four of which also include ISA connectors. Even the add-in cards require no tools when you replace a card. Instead, hinged plastic restraints lock boards into place. I swapped PCI graphics boards in less than 30 seconds. The system provides equally easy access to peripherals. Unscrewing a couple of thumbscrews lets you open a hinge-mounted front panel, after which you can easily add or change disk drives or other peripherals. The system sports 10 drive bays with front access and a separate hard-drive cage that can hold four to six disk drives, depending on whether you use 1- or 1.6-in. drives. In the rare case that you need better access, unscrewing a few thumbscrews allows you to remove the side or top sheet-metal skin.

In ease of access, only Dell's new workstation family Personal Workstation 400M, with prices starting at $3899, comes close to matching the Compaq system. And, although I didn't get a chance to test the Dell system, I saw a sample unit at a trade show. With the Dell system, you must first snap off the sheet-metal skin on the left side of the tower case. You then have easy access to the add-in cards and memory. You can move a hinged power supply from the enclosure, leaving plenty of work space. The Dell system, however, appears to have the least robust cooling system of all the Wintel workstations that I've seen--a characteristic that probably results from the fact that Dell developed the chassis for its high-end PC line.

As with the Dell system, all of the other systems I investigated require that you remove a skin or the outer cover for access to memory, µPs, add-in cards, or peripherals. You can open all these covers without tools, but you still have to remove the cover and work in tight quarters. I tested two Intergraph systems, TD-325 and -425, which come in slim, pizza-box-style packages with little spare space. Intergraph designed the systems, however, so that you can remove items, such as the floppy-disk drives, without tools. It lets you remove enough items in a couple of minutes to easily access µPs and memory. Intergraph has introduced a tower system, the TDZ-2000, that the company labels as "tool-less," which uses hinged assemblies to simplify access. It also features an optional snap-on base that extends PCI/ISA-card capacity from eight to 13 slots and disk-drive bays from six to 11.

Digital has also developed a unique package in its Model 600 au series of workstations. You unscrew two thumbscrews to remove the skin, yielding easy access to DIMM sockets and PCI slots. Assuming that you've removed all PCI cards, the design also allows you to slide out the motherboard for replacement. And workstation customers are accustomed to the concept of motherboard upgrades, whereas PC users have been more likely to buy an entire new system.

The new HP Kayak system proves to be the toughest system to work in but may just be the best in airflow and rugged mechanics. The tower design is smaller than the Compaq system and therefore has less free space. To replace a disk drive or access the four DIMM sockets, you must remove the cover, loosen thumbscrews, partially remove the power supply, and remove a cooling fan/air-duct assembly. The cooling assembly, however, indicates a robust mechanical design with temperature-controlled fans; the rigid plastic air-control ducting is unique among the systems. HP claims that the UltraFlow cooling scheme directly cools hard drives and PCI/Accelerated Graphics Port (AGP) slots as well as the µPs. Moreover, the reinforced chassis in the HP system would survive greater shock than the other systems.

At least the mechanical comparison is relatively straightforward. I thought that a performance comparison would be straightforward as well, but my evaluation of the three systems convinced me otherwise. I could have turned to standard PC benchmarks, such as Winstone or Winmark, to test performance, but I didn't think that those test would yield useful data for EDA users. The tests are designed for Windows 95, whereas most EDA users will probably use Windows NT (see box "Windows 95 on workstations"). Moreover, those tests are based on office-productivity applications that emphasize interactive document changes and screen updates. Meanwhile, EDA users would likely be more interested in CPU-intensive tests, such as autorouter or Spice-simulation algorithms.

I used EDA software to test the systems. I concluded that routing a sample pc board would be a test that should be familiar to most designers. I contacted Accel Technologies, which provided me with a copy of the Accel EDA software suite. A few years ago, Accel purchased P-CAD, and the new Accel EDA Version 13 represents the merged offerings of both companies. The tool suite includes a number of modular tools based on the $995 Accel Schematic schematic-capture program and the $7950 Accel P-CAD PCB pc-board-layout package. A version of the pc-board package that supports boards with six or fewer layers sells for $3995.

Accel offers two autorouter options with its EDA suite. The Accel ProRoute package costs $5500, and a version limited to four signal layers and two power layers costs $2950. Accel licensed ProRoute from Intergraph's EDA Division, which is now VeriBest (Boulder, CO). ProRoute is the lone portion of Accel EDA that isn't a true 32-bit Windows application. Accel also offers the $9000 shape-based Spectra 7.1 router that it licenses from Cadence. You can buy a version of Spectra that's limited to four layers for $2300. Because of ProRoute's simpler licensing and installation factors, I decided to use the ProRoute software and test each system in single- and dual- processor configurations with 64, 128, and 256 Mbytes of memory.

The three systems provide a good cross-section of what vendors offer. All support one or two 266-MHz Pentium II µPs, and I also tested the Kayak with 300-MHz µPs. Intergraph introduced the one-processor TD-325 and the two-processor TD-425 early this year, and they use older Intel FX core logic. Compaq introduced the Professional in July. It uses a chip set from Reliance Computer Corp (RCC) that essentially offers a crossbar architecture in which both µPs share access to two PCI buses and two memory banks. The HP Kayak uses the new Intel LX chip set with support for the AGP.

Other differences could also affect performance. For example, both the Intergraph and the Compaq systems use extended-data-out (EDO) DRAM, whereas the Kayak uses faster synchronous DRAM with error-correction code (ECC). Intergraph's new TDZ-2000 will also use ECC-based synchronous DRAM. All of the systems use Ultra fast and wide SCSI disk drives, although the HP system also comes with a RAID option.

All of the systems also use the Matrox (Dorval, PQ, Canada) Millennium graphics board. The Intergraph system uses the standard Millennium, the Compaq uses the Millennium II, and the HP uses the AGP version of the Millennium II. However, the predominantly 2-D EDA applications make the graphics board a nonfactor in performance comparisons. On the other hand, all of these vendors offer an array of 3-D graphics boards that can challenge Silicon Graphics (Mountain View, CA) systems. If your application requires 3-D graphics, evaluate 3-D architectures before considering the overall workstation architectures.

I ran a test to sample the relative capabilities of these machines, computing the amount of time that each system requires to route a typical 386-era motherboard with four signal layers and power and ground planes (Table 1). The table also lists the same test on a Windows 95 Pentium-class system with a Cyrix (Richardson, TX) P166+ µP and 32 Mbytes of DRAM.

Early in the testing, I recognized that my results were sometimes exceeding my expectations and were sometimes not even near my expectations. I first ran the tests on the Compaq system and was pleasantly surprised that the dual-processor configurations provide a significant speed boost: My first set of tests revealed more than double performance with dual µPs, even though I was using an application that had not been designed to take advantage of two µPs. I was using the multiprocessor Windows NT kernel, so I expected the second processor to help because the OS functions can run on one µP while the application runs on the second µP.

Although I was somewhat new to NT, I suspected a problem with the OS configuration in the one-processor test. The beta unit I had received from Compaq came with instructions for switching between one and two processors. The instructions told me how to add or remove a processor pc board but said nothing about how to change the OS. It turns out that running the multiprocessing NT kernel on a single-processor system works but adds significant overhead. Most users, however, would rarely start with a two-processor system and remove one, so it's doubtful that you would encounter the same situation. It's easy to make Windows NT run properly for one or two processors. You can create two boot configurations and select the normal or multiprocessor mode at boot time. The difference between the two boot configurations comes down to two files--hal.dll and ntosknl.exe--that are typically in the system32 subdirectory of Windows NT.

In some EDA applications, meanwhile, the dual processors still offer a 30 to 50% performance boost, even on software that's unaware of the second processor. I generally found a similar advantage running office-productivity applications. Moreover, dual processors provide a significant advantage when you simultaneously run multiple applications. I would have liked to run some multiprocessing EDA tools, but few if any are available. However, multiprocessing software for animation and imaging applications does exist, and perhaps next year we can revisit NT workstations for a look at multiprocessing EDA software.

The first set of tests revealed an unexpected trend when I varied the memory configurations. I had first run the test on the Intergraph and Compaq systems, and in both cases the 64-Mbyte configurations resulted in better performance than the 128- or 256-Mbyte configurations. I wouldn't have been surprised had the tests revealed no difference in performance because I wasn't sure that my sample pc-board-route application needed more than 64 Mbytes of memory. The test, however, ran significantly faster with 64 Mbytes of memory.

I turned to Accel Technologies to find out why I got the results I did. Although, the company had just upgraded its new Version 13 package to the 32-bit Windows model, I chose the one portion of the Accel EDA suite--the autorouter--that Accel licenses from other companies. ProRoute uses a routing algorithm that essentially runs in an extended MS-DOS mode. The router recognizes no more than 64 Mbytes of memory and suffers random confusion when you run it on a system with more memory. At this point, I began to suspect that you had better choose your software before your hardware to get the best performance from an NT workstation. By choosing software first, you can directly compare systems and configurations. You can never predict when you might hit a legacy-code issue.

By this point in my testing, I still hadn't received the HP Kayak, so I found a last-minute alternative to ProRoute that might show better performance with large amounts of memory. Accel Technologies also licenses the Spectra shape-based router from Cadence, so I obtained a 14-day evaluation license that I could use to test the systems. The Accel tech-support staff assured me that Spectra 7.1 was a 32-bit application that could use lots of memory.

I began to repeat the test on the Compaq and Intergraph systems, although again I found some surprising results: The router ran best with 64 Mbytes of memory, although the differences were small. Windows NT suffers from some memory-management overhead. Therefore, slightly degraded performance occurs when your system has more memory than it needs. I considered trying to route a more complex pc board that could use more memory, but deadline pressures were mounting. Moreover, I wasn't sure I had time to run such a test in a variety of processor and memory configurations on multiple machines. Also, a friend that uses Spectra told me that an appropriate test might take hours rather than minutes to run on each configuration.

Using Spectra also revealed the disadvantage of paying for a second processor in some cases: The Spectra router spends much of its time in computation-intensive geometry routines and little time making OS calls. Therefore, the second processor provides little advantage and, in some cases, slightly degrades performance, probably because of the overhead in the multiprocessing version of the NT kernel. For an application such as Spectra, you would be better off with more expensive disk and memory systems and one processor. On the other hand, a second processor would let you run a second or third application with little impact on Spectra.

After running both the Spectra and the ProRoute tests on the Intergraph and Compaq systems, I finally got my hands on an HP Kayak. The system promised to be a good platform to test the merits of a single processor with advanced disk and memory systems. The Kayak I received had a dual-drive HP FastRAID and two 10,000-rpm SCSI drives from Seagate (Scotts Valley, CA). The RAID option implements both RAID-0 striping and hardware caching to the fast and wide Ultra SCSI bus. I also tested the Kayak with a single Ultra SCSI wide and fast drive without caching. Moreover, HP supplied a 300-MHz system that I could also slow to 266 MHz for comparison with the other systems.

The RAID controller delivered a consistent but relatively small advantage in my tests (Table 1). The advantage of RAID would be more apparent in I/O-intensive applications, but such applications are relatively unusual in EDA. You would also see important benefits if your applications demand more memory than your system has and the OS spends significant time paging memory to the disk subsystem. Tests with the Kayak also provide a clue to what you might expect from incremental changes in CPU-clock rate. Even a slightly faster processor provides more improvement in performance than do RAID schemes.

Again, however, tests with a new hardware-and-software combination produced some strange results. The Kayak struggled to run the ProRoute test much faster than my Pentium-class reference system could. I had to track down the Kayak designers in France to solve this mystery. It turns out that BIOS designers must work magic to maintain legacy compatibility with DOS applications, and such compatibility is especially important in Windows 95. DOS applications use an Interrupt 15 call to discern how much memory is available in a system, but that software function supports only as much as 64 Mbytes of memory. Windows 95 and NT systems support a different function call for 32- and even 16-bit Windows applications.

The HP BIOS returns an accurate memory, accounting for systems with 48 Mbytes of memory or less. However, the BIOS designers decided to report only 16 Mbytes of free memory in a 64-Mbyte configuration to ensure that all legacy software would run on the 64-Mbyte system. This phenomenon effectively limits memory and causes the Kayak to always take 23 minutes or more to run the ProRoute test--yet another indication of the importance of choosing software first.

Despite the surprises I encountered with these workstations, I still believe they represent the future for EDA applications. And, despite the lack of multiprocessing software, you can theoretically evaluate the merits of the architectures. Compaq has taken a possibly risky step in choosing a multiprocessor chip set from a company other than Intel, but it's a step that may pay off. The RCC chip set employs two dedicated memory controllers for the dual memory banks, and the dual PCI buses could offer an advantage in a multiprocessor environment. Compaq claims that the memory buses provide aggregate bandwidth of 1.07 Gbytes/sec--a number more akin to system architectures such as Sun's crossbar-based UltraSPARC Port Architecture than to high-end PCs.

Let's hope that by next year EDA tool vendors will have multiprocessing-enabled applications that you can use not only to test NT workstations but also to take advantage of the NT systems. One of the great promises of Wintel workstations is that even with multiple µPs, the price stays low, because PC volumes make the processors relatively inexpensive. Most silicon vendors say that it's easier to shrink and lower cost on a current µP design than it is to double clock rates. Unfortunately, the multiprocessor approach depends on rebuilt applications that can truly use the available resources.

Digital, for one, doesn't buy the dual-processor argument. The company will sell multiprocessor systems but is also constantly pushing clock speeds in the Alpha µP. Digital's newest systems may provide the biggest challenge to Wintel workstations, although Digital also sells Wintel machines. The Alpha µP is the only processor that has steered clear of the Intel juggernaut, although rumors persist as this article goes to press that Intel would buy Digital's Alpha technology. Digital's Alpha-based personal-workstation line now tops out with a 600-MHz, 64-bit µP, yet the company claims it brings PC economics in tow. The system uses the PCI bus and standard peripherals. Prices start at approximately $10,000.

Because Digital also sells Wintel machines, the company is careful about comparing Alpha and Wintel performance. Some of Digital's Alpha partners, however, claim that the machines outperform Pentium II systems running Intel code on the FX32 simulator. Undoubtedly, however, you could face legacy-code problems, so you probably shouldn't buy an Alpha system to run Intel code. If you find your chosen applications in an Alpha version, however, the technology offers a viable choice, and you can still use FX32 to run the occasional Wintel application. Moreover, you have the advantage of still using NT, which seems destined to control the desktop.

Other than Alpha, only Sun and Sun-clone vendors seem set to resist Intel. As this issue goes to press, Sun has announced a 600-MHz UltraSPARC III that it claims will significantly outperform the Pentium II. Sun claims it will ship the µP in mid-1998. Still, Sun may regret its decision down the road. Digital took many years to produce 500-MHz Alpha µPs with cost-effective yields. Any delay will mean that UltraSPARC III must compete with the next-generation Merced µP, which HP and Intel developed, rather than the Pentium II. Even Silicon Graphics plans a Wintel system, and that could signal the beginning of the end for MIPS (Mountain View, CA) µPs in workstations and servers. Rumors also persist that eroding market share and lack of support from Apple's (Cupertino, CA) silicon partners could force the company away from the PowerPC.

HP understood Intel's momentum several years ago when it signed on to cooperate and jointly develop the Merced µP. That µP will allow HP to move to a workstation-hardware base yet offer a choice of NT or Unix. Intel and HP developed Merced in a way that supposedly will allow it to run Intel or PA-RISC binary code. All along, HP has claimed that it would continue to use the PA-RISC code it has supported for years. Now, however, HP indicates that Merced will assure its users a smooth transition for PA-RISC-legacy applications. However, HP admits that a future transition to Intel code is inevitable.

Use this article as a road map when you evaluate Wintel workstations, but don't limit your choices to those systems mentioned. Vendors such as Micron (Nampa, ID) are entering the market, and companies such as Intergraph and Compaq have far more systems available than those mentioned. For example, Compaq also offers a four-processor PentiumPro system that could serve in some applications, and a new dual Pentium II desktop model called the 5100 series. Intel is pushing the PentiumPro for servers and the multimedia-extension-equipped Pentium II for workstations. The Pentium II design effectively limits systems to two processors. If your application can take advantage of four µPs, the slightly older technology can be a better choice.

Reference

1. Wright, Maury, "Workstation vendors brace for Pentium Pro whirlwind," EDN, Sept 2, 1996, pg 59.

| EDN Access | Feedback | Table of Contents |