![](/ProductImages/0507/pd0507airplane.jpg") | ![](/ProductImages/0507/pd0507helicopter.jpg") |
![](/ProductImages/0507/pd0507truck.jpg") |
With 64-bit processing, engineers using MSC Virtual Product Development
software can simulate and analyze scenarios such as airplane crashes. |
Both 64-bit processors and supercomputers are poised to make in-depth component modeling and enhanced simulation possible for design engineers.
By Bart Eisenberg, Technical Editor
Will advances in hardware change the way you work? The companies that make CPUs and operating systems certainly want you to think so, as do many, but not all, CAE vendors. For engineers looking to the horizon, two technologies — 64-bit processors and cheaper supercomputers — are working their way from opposite ends of the computing spectrum toward broader adoption. For both technologies, design engineers are among the most prized early adopters.
Fueled by the need to sell ever more CPUs to customers who already own them, Intel and especially AMD have pushed 64-bit as the next big thing in computing, the technology that will transform the ordinary Windows-based PC into a full-blown workstation. Sometimes, the chipmakers get ahead of the curve. AMD published an effusive account on Bell Helicopter's purchase of AMD Opteron-based 64-bit workstations. But none of the speed advantages were due to 64-bit applications because the group was still running the 32-bit version of Windows.
The main advantage of the 64-bit architecture is its ability to address more memory. A 32-bit CAE application running on a 32-bit processor can address up to 4 gigabytes of memory, typically giving Windows applications up to 2 GB of usable RAM on a properly configured PC. The same is true if that application runs unaltered on a 64-bit processor like the AMD Opteron, Intel Xeon, or Intel Pentium 4. But if that application is rewritten to support these processors' 64-bit mode, more than a terabyte (1,028 GB) of memory can be addressed — at least in theory. The primary restriction is the operating system, which must support the processor's 64-bit instruction set.
Two Microsoft products, Windows XP Professional x64 Edition and, further out, Longhorn, are poised to become the 64-bit operating systems of choice. The new version of XP will initially support up to 128 GB of RAM, which is still in the stratosphere for a desktop machine. Applications can further limit this amount. MSC.Nastran, for example, requires contiguous memory, and its solver, the number crunching portion of the application, supports up to 8 GB.
Analysis and Advantages
Analysis and simulation are likely to be among the first beneficiaries of 64-bit computing. John Buchowski, director of product management at PTC, says that the memory limitations of 32-bit has required simplified models, which takes expertise to produce. "You need to know whether or not the detail you got rid of was important. With memory not being a limitation, I can run my models exactly as they were designed." One PTC customer is using 64-bit computing to test transmissions, “an extremely complicated casting, with lots of ribs, fillets, and holes, and it's typical of what can cause a lot of complexity for FEA code," Buchowski says.
Some “advantages” of 64-bit are misunderstood. "A common misperception is that 64-bit affects accuracy," says Kevin Kilroy, who heads testing and porting for MSC.Nastran. On MSC.Nastran, all calculations are done with 64-bit, double precision floating point, no matter what the architecture. Nor in most cases, does 64-bit computing handle more complex analysis than 32-bit. The biggest advantage of 64-bit application analysis is speed. "It's problem-dependent," Kilroy says. "You have to have a problem that needs the memory space, and until you get to that limit, there is no difference."
To answer a reporter's query into what size problem might benefit from 64-bit, Kilroy and his colleagues ran a benchmark test on a static solid model. With up to 500,000 degrees of freedom, 64-bit had no speed advantage over 32-bit. But as the df count approached a million, throughput nearly doubled. MSC.Nastran's solver’s contiguous memory requirement means that those first 8 GB of memory make the biggest difference. With yet more memory, the application still runs faster because fewer disk operations are required — though the improvement is less dramatic: 5 to 18 percent on benchmark tests. While actual results will vary, Kilroy suggests this is a rough but reasonable rule-of-thumb.
The added speed can benefit the analysis of many kinds of models — even a simple bracket. "The model doesn't have to be complex if you make the mesh fine enough," Kilroy says. "In the past, you would make the mesh finer around the bolt hole or other area of interest, then coarser further out. With more memory, you could use a fine mesh throughout." For a more complex structure like an engine block, the advantages are obvious.
ADVERTISEMENT
What about CAD?
The advantages of 64-bit CAD are less obvious, applying primarily to large-component design projects like aircraft and automobiles. "But the sweet spot for our customers is in machine design, such as the line in at a bottling plant," says Fielder Hiss, who heads product management at SolidWorks Corp. "A single capping maching might have 10,000 to 15,000 components. While 32-bit CAD might well handle one such machine, it loses steam trying to model several machines that comprise a complete assembly line.
One current work-around already built into the company's design packages is to limit the data in order to model more components. Graphical information is displayed, but not necessarily the underlying engineering data, such as mass and center of gravity. A 64-bit application would accommodate the complete model, rather than just a subset. "The more data you load into the system, the more memory is required," says Hiss.
SolidWorks, whose products run exclusively on Microsoft Windows, has not at this writing announced 64-bit support. PTC, on the other hand, has a 64-bit version of Pro/Engineer which it is porting to workstations running Windows XP Professional x64 Edition. The company already supports various flavors of Unix, which account for about 15 percent of its installed base. So if there is to be a tipping point for 64-bit, Windows will be it. Potential applications include the ability to model the interaction between equipment on a drilling rig. With conventional 32-bit, the outer surfaces could have been represented, but not the underlying components.
Supercomputing’s Future
While 64-bit represents a comparatively modest transition of standard-issue PCs and servers, high performance computing (HPC) is a bigger revolution and a far more costly one. HPC makes use of a technique called parallel processing, in which pieces of a single application are processed at the same time. The most popular way to do this is with a clustering: connect several computers so that they behave as one. Here, the Linix operating system is pervasive, running on approximately 60 percent of the world's top 500 supercomputers. HPC has moved from research labs to the aeronautics and automobile industries, and with prices falling, is poised to reach beyond.
Dave Turek, vice president of "deep computing" for IBM's systems and technology group, says that as the size of clustered systems shrinks, these machines will fit in more places. "NEC's Earth Simulator, which from 2002-2004 was the fastest supercomputer in the world, required more than a football field to house it," he says. "The electrical bill to run and cool it ran in the millions of dollars annually." Last fall, the IBM Blue Jean/L system surpassed the NEC not only in terms of compute power — 70.7 teraflops versus 35.8 — but also in low power consumption. With additional processors, it has since delivered 135.5 teraflops. And the system is much smaller than the NEC and able to fit in a large living room.
IBM is now developing a 360 teraflop BlueGene IBM for Lawrence Livermore National Labs that will consume just 1.5 megawatts. The secret is substituting low-power microprocessors designed for embedded applications with conventional CPUs. These processors are also integrating more functions, thereby doing more in a smaller space. One classic laboratory application for HPC clusters is in atmospheric forecasting — and so fluid dynamics may well be the intersection where academic research meets mechanical engineering.
Turek says that a new generation of clustered systems is beginning to support today's applications. "You can now deploy them in work environments that are less restrictive, fitting a multi-thousand node cluster in a conventional rack mount that delivers almost six teraflops of computing." This compute power will come at ever shrinking prices. These aren't hobbyist systems. Turek puts the entry-level cost at $2 million.
Simulation Savings
But if the simulations can reduce the number of physical test, increasing the number and variety of simulations, design companies beyond automotive/aerospace will find themselves the proud owner of a supercomputing cluster. Stan Posey, field business developer manager for SGI, says that the real cost of HPC is in software and U.S. engineering salaries, which are outpacing the cost of hardware. “Cheaper processors are making it possible to get more for your money in a single system. Engineers still tend to use 12-16 processors for single CAE simulation; we've shown the advantages of using 128." But the cost of software can be prohibitive for a single user who wants to deploy 128 processors for a single simulation. That’s starting to change with more affordable “parallel” licensing fees and as engineers find ways to cut costs by conducting more simulation.
Those savings come not only from fewer physical tests, but also from the number of iterations an engineering group can afford to conduct. For example, says Posey, consider a gas turbine engine housing designed to confine the damage from a blade failure. Engineers can inexpensively run simulated envelope studies that consider a large number of impact orientations and engine designs. By contrast, the cost of changing even a few parameters through trial and error physical testing are much higher. A succession of laboratory measurements have become ever more expensive because of rising labor costs and precision instrumentation.
Posey's turbine engine example demonstrates another trend with HPC. While the aerospace and automotive companies remain the principal users of CAE technology, HPC is beginning to trickle down to the component makers. If the economics work out, HPC simulations for subassembly designs will become a mandate.
For product designers looking ahead, the two technologies will play out differently. HPC will be an expensive option with obvious benefits for at least some design groups. They will adapt it if and when the cost versus benefits ratio makes sense. But sooner or later 64-bit will arrive on desktops because the chip-makers and Microsoft see it as their future. How the CAE application companies leverage the architecture remains an open question. If history tells us anything, only when a hardware advance becomes widespread do the software developers discover how to take full advantage of it.
NEXT STEPMore information on AMD is available at www.amd.com or by calling 602-242-4400.More information on IBM is available at www.ibm.com or by calling 800-426-4968.More information on Intel is available at www.intel.com or by calling 800-538-3373.More information on MSC.Nastran is available at www.mscsoftware.com or by calling 714-540-8900.More information on PTC is available at www.ptc.com or by calling 781-370-5000.More information on SGI is available at www.sgi.com or by calling 650-960-1980.More information on SolidWorks Corp. is available at www.solidworks.com or by calling 800-693-9000.