Moore’s law says computer processing speed will double every 18 months. So that the law holds into the foreseeable future, CPU and computer builders have been working overtime to keep new systems in the pipeline. For example, Intel CPUs have climbed to 450 MHz and show no sign of leveling out. Computer builders are adding several buses on motherboards to keep processors fed and data flowing to several subsystems at once. Most NT and a few Unix workstations allow adding a second processor. And equally impressive is that for about $5,000 on the low end, dual-CPU computers can be had for solving tough problems faster than previously possible.
Spreading one problem over two CPUs to shorten solution time is the most often cited advantage for the machines. There are others. For example, Research from Ansys Inc., Pittsburgh, shows that even when running a nonthreaded program (one not written for running on two processors) solution time can be cut by as much as 30% because operating system functions are handled by one CPU and application functions on the other. You might also use the machines by kicking off an analysis or a long print job on one CPU and start modeling in solids on the other. Several networking schemes accomplish similar feats on a larger scale by ganging idle CPUs into a distributed multiprocessor computer for solving time-consuming problems after hours. One SGI workstation makes good use of dual CPUs by letting two users perform separate tasks on the same computer.
It’s not surprising that single-processor computers far outnumber dual-CPU designs. Two-CPU computers cost more and there’s limited engineering software to take advantage of tandem processors. A few FEA programs that have been multithreaded include Ansys from Ansys Inc., Pittsburgh, and MSC/Nastran from MacNeal-Schwendler Corp., Los Angeles. Many CFD programs, such as Fluent from Fluent Inc., Lebanon, N.H., are written with threads. Further, rendering software and those aimed at the animation industry are multithreaded.
Windows NT and most versions of Unix automatically handle a few parallel functions. Tests with the Ansys FEA program show that multithreading slashes up to 46% off the run time needed by a single-processor computer. But don’t expect to lop half the time off every solution. Software vendors agree that the amount of compute time trimmed depends on the application.
Despite the shortage of software, Intergraph reports that about 15% of their sales are dual-processor machines, and there’s no telling how many single-processor machines have been upgraded to dual status. Admittedly, if your work centers around 2D drawings, the two-CPU hardware might be less useful. But as models and assemblies grow larger and rebuild times extend beyond coffee breaks, the technology makes more sense.
Xeon — the latest Pentium
At the heart of most new dual-processor computers will be the latest Pentium II or Xeon processor. It’s aimed squarely at compute-intensive engineering functions and packs in several improvements made over the last generation. Intel is calling it a Slot II processor because its 330 pins (versus 240 pins on the previous Pentium II) require a larger connector. The company will continue production of high-clock speed Slot I processors.
The first Xeon design ran at 400 MHz and 450-MHz versions are coming on-line. Cache speed and size have been reworked. For instance, first and second levels of cache now work at full clock speed. Level one on the previous design accessed data at only half the clock speed. The development is significant because cache provides data for immediate processor operations. When the cache stays full, the processor spends little time idling.
Next to the 32-bit Xeon, the most significant development in Pentium-based workstations is the AGP or accelerated graphic port. It’s more than a connector on the back as its name suggests. It’s also a bus that speeds graphic data to the monitor. A 440GX chip set sits at the center of the data-flow system. (The Xeon gets a 440 GX chip set while a 440BX or LX appears in previous Pentium computers.)
“The 440GX is the AGP hardware that connects to the PCI bus, provides the PCI architecture for peripherals, and controls data flow between memory and CPU,” says Kathelene Tandy, a computer specialist with Hewlett-Packard Co., Boston. “The BX chipsets on non-Xeon-based computers control the front-side bus (between CPU and main memory) at 100 MHz. The GX chips work with Xeon processors and support up to 2 Gbytes of RAM versus 1 Gbyte for the BX,” she says.
Even the AGP bus is evolving fast. The design operates at 66 MHz and moves 4 bytes/clock cycle or 267 Mbytes/sec. The next version that’s on most Intel-based computers mentioned in this article runs at 534 Mbytes/sec. Watch for the AGP Pro in upcoming 450-MHz systems says Barry Luck, a development engineer with IBM’s IntelliStation group in Raleigh. “It’ll have a bandwidth of about 1,068 Mbytes/sec.” This next-generation AGP will give additional power to graphic cards.
To round out many of the Intel-based systems, Ultra SCSI hard drives will be running close to 10,000 rpm and capable of spitting out data at 80 Mbytes/sec.
A glance at the available systems might suggest that they’re all alike. That’s not true. Companies such as Sun, HP, Compaq, Intergraph, SGI, and IBM build their own motherboards which lets them tweak their systems for performance advantages. For another perspective on performance, check out www.specbench.org/gpc/opc/opc.cdrs.summary.performance.html. The accompanying boxes detail a few systems configurations.
Ultra 60 tuned to 3D modeling and prototyping
The UltraSparc II processor and Ultra Port Architecture crossbar technology supports a 64-bit, 66-MHz PCI I/O bus. The computer has double the components and double the capacity for better overall performance and greater flexibility. In addition to the two 360-MHz processors, the system also sports two disk bays, two UPA (Ultra Port Architecture) accelerated graphic ports, two PCI channels, two Ultra SCSI channels (internal and external), 2-Mbytes external cache, and 2 Gbytes of memory. CPU data paths are 144-bits wide with 128 bits for data and 16 bits for EDD. The Ultra 60 runs the UPA at 120 MHz (up to 1.9 Gbytes/sec throughput) to support the processors.