It's hard to overestimate the importance of Advanced Micro Devices Inc.'s Opteron x86, 64-bit processor. In just over a year, the three major enterprise hardware vendors have built offerings around it, and AMD's biggest competitor, Intel Corp., is producing its own version of the chip. Opteron is forcing change.
What this means for users is this: The 32-bit-only processor is about to go the way of the 16-bit chip. "By year-end, we will be selling very, very few 32-bit systems," said Paul Miller, vice president for industry-standard servers at Hewlett-Packard Co.
As users refresh or add industry-standard servers, they'll seed their data centers with 64-bit-capable x86 processors. The pricing differences between current 32-bit boxes and those running on 64-bit chips will be negligible in the near future, so the reason for buying 32-bit chips will gradually disappear, say vendors and analysts. Think of 64-bit capability as a free upgrade, ready for use when the applications arrive.
Opteron has gained a lot of attention because it allows users to run 32-bit and 64-bit x86 applications on the same chip, giving users the flexibility to gradually move to the 64-bit world. But the chip also includes architectural changes that may speed a 32-bit application's performance.
For now, early Opteron adopters like Aristotle Balogh, senior vice president of operations and infrastructure at VeriSign Inc. in Mountain View, Calif., tend to have memory-hungry custom-built applications. VeriSign already uses many RISC- and Intel-based servers to support, among other things, directory services for the .com registry. It has been testing two- and four-processor Opteron systems, putting tremendous processing loads on them in a "beat the box up until it drops" test, and the chip has performed well, Balogh says.
Balogh can get 64-bit capacity from RISC-based Unix systems. But a four-processor Opteron box with 32GB of memory will cost about $25,000, whereas a Unix box may cost more than $100,000. "With traditional Unix vendors, it is a very expensive proposition," says Balogh.
The memory gain allowed by a 64-bit chip is a big advantage, but it's not the only one. Opteron, which can run 32- and 64-bit applications, is gathering support from some 32-bit users because of how it's designed.
AMD has developed what it calls HyperTransport technology, which directly connects the CPU to the memory, eliminating the need for a bus. This reduces latency and speeds processing time, which is why Automated Trading Desk LLC, a company that provides trading technology and financial trading services, started using Opteron on Altus servers from San Francisco-based Penguin Computing Inc.
Eric Hunter, senior Linux systems administrator at the Mount Pleasant, S.C.-based company, says it runs custom-built applications that use a lot of memory, and "getting rid of the bottleneck" between the memory and CPU was the main concern. "We just saw tremendous increases in performance in our test box," he says.
Sixty-four-bit processing power is suited to programs that require large data sets that need to go above a 32-bit processor's 4GB memory limitation. Many of the applications taking advantage of this today are scientific and involve design and rendering.
For instance, the University of Utah's Center for High-Performance Computing in Salt Lake City has been recompiling its 32-bit applications to 64 bits, using a compiler developed by PathScale Inc. in Sunnyvale, Calif. This speeds up processing time by 10% to 20%, depending on the application—a significant gain for applications that run over many hours, said Martin Cuma, who is in charge of scientific application programming at the center.
But most important for the university is the increased memory addressability. Sixty-four-bit computing has enormous memory potential, calculated two to the 64th power—many terabytes. Instead of simulations that have 100 atoms, for instance, researchers can run them with 200 atoms. "All of these applications are really memory-hungry," says Cuma.
Applications Needed
For users that don't have an immediate need for those memory gains, the arrival of 64-bit applications will drive adoption, says Don McPherson, network operating systems and database administrator for a nonprofit organization that he asked not be identified.
McPherson says he can see a need for 64-bit memory because of the demands being imposed on databases. "We're pushing more data, and doing more things with databases," he says.
Analysts, vendors and users aren't certain how long it will take 64-bit computing to become mainstream. But operating systems that support x86-based 64-bit chips are arriving. There are 64-bit versions of Linux already available, and Microsoft intends to release a 64-bit version of Windows XP by the end of the year. AMD and Intel are expected to be binary-compatible with Windows 64-bit.
"We expect the transition to 64-bit software will be fairly slow," says Jon Sharp, director of platform marketing at Intel, which plans to release its own x86, 64-bit chip this summer, the EM64T. "The transition to 64-bit operating systems will happen somewhat faster," he says. "What people will care about is being able to run the two applications, 32-bit and 64-bit, side by side."
Intel views its Itanium 64-bit chip, which uses a different architecture from x86, as a challenger to the high-end RISC systems. The chip has "massively parallel resources," larger cache and more memory bandwidth, says Sharp. For instance, Itanium is well suited for use in large, multiprocessor scale-up systems, which have been Unix's domain, he says.
It's not a question of if people will go to 64 bits. "It's a question of when," says Paul Terry, chief technical officer at Cray Canada Inc., whose parent company, Cray Inc., is building a 10,000-processor Opteron system for the U.S. Department of Energy's Sandia National Laboratories.
|