64-Bit Servers

When Intel introduced the 64-bit Itanium processor in 2001, most buyers (and many server vendors) virtually ignored the chip because of its poor performance and weak third-party software support. If you haven't been following this segment of the server market during the past 2 years, you might want to take a second look; powerful new processors from AMD and Intel make migrating to 64-bit servers more compelling than ever.

Last year, Intel introduced the second-generation Itanium 2 processor (code-named Madison), which, according to published Transaction Processing Performance Council TPC-C test results (http://www.tpc.org/tpcc/results/tpcc_result_detail.asp?id=103082701), delivers on the 64-bit performance promises of its Explicitly Parallel Instruction Computing (EPIC) architecture. As a result, enterprise software and server vendors are now supporting the 64-bit processor. In addition, AMD released its long-awaited 64-bit AMD Opteron processor last year, and Microsoft has promised to support the chip with a soon-to-be-released 64-bit version of Windows Server 2003.

Because Microsoft hasn't yet released a version of 64-bit Windows for AMD's chip, comparing the 64-bit performance of Itanium 2 with Opteron under Windows is impossible. But published TPC-C test results (http://www.tpc.org/tpcc/results/tpcc_result_detail.asp?id=10309091) demonstrate that the Opteron's performance using current 32-bit applications and Windows 2003 is extremely good. Because 32-bit performance has been Itanium's Achilles heel, Opteron could address a major concern of 64-bit server buyers. If you aren't ready to migrate your 32-bit applications to their 64-bit versions (or can't because 64-bit versions aren't available), your server applications are bumping up against the 32-bit system's 4GB memory limits, or your applications need higher performance levels than current Intel Xeon-based servers can provide, now might be the time to reevaluate your server strategy.

The Processors
Intel 64-bit processors debuted in 2001, with the introduction of Intel's first-generation 733MHz and 800MHz Itanium processors (code-named Merced). The first Itanium 2 processors (code-named McKinley) were 900MHz and 1GHz models that featured an improved architecture that substantially improved performance while maintaining binary compatibility with the original Itanium. Madison, the current generation of Itanium 2, uses a smaller process size (i.e., the average size of a chip's features) resulting in shorter electrical paths and a smaller surface area. The new manufacturing process lets Intel increase Madison's clock speeds to 1.3GHz, 1.4GHz, and 1.5GHz and enlarge L3 cache sizes to 3MB, 4MB, and 6MB, respectively, from McKinley's 1.5MB and 3MB cache sizes. Table 1, page 24, compares McKinley's and Madison's technical specifications with those of the original Itanium (Merced).

At press time, Intel introduced a lower-priced 1.4GHz version of Madison for dual-processor applications; this model features 1.5MB of L3 cache. The company also introduced a low-voltage 1GHz version of Madison with 1.5MB of L3 cache; the company claims that this processor uses about 50 percent less power than other Itanium 2 models. Intel's 8870 core logic chipset supports 2 to 16 processors; however, some first-tier server vendors have developed their own core logic chipsets that support more than 16 processors and offer additional capabilities.

AMD has released 1.4GHz, 1.6GHz, 1.8GHz, and 2GHz versions of its 64-bit Opteron processor. All models are available in 100, 200, and 800 series versions that support servers with one, two, or as many as eight processors, respectively, when paired with the AMD-8000 series core logic chipset. (Hardware vendors can use third-party chipsets to build servers that have more than eight processors.) Table 2 lists the Opteron's technical specifications. Don't assume that Operon's higher clock speeds necessarily translate into performance better than Itanium 2's. The different instruction sets and architectures make clock-speed comparisons meaningless. As you can see from Table 2 and Table 3, page 26, the Opteron's 1000-unit pricing for 100 and 200 series processors is substantially less than the Itanium 2's pricing and should help system vendors create 64-bit Opteron-based servers that cost less than similarly configured Itanium 2–based systems.

AMD and Intel have taken different approaches to 64-bit computing. The Itanium 2 employs an entirely new architecture and instruction set, whereas the Opteron extends the existing x86 architecture. Both vendors claim 64-bit performance that substantially exceeds the performance of similarly configured Xeon-based servers.

32-Bit Compatibility. Unlike the Itanium 2, the Opteron maintains full binary compatibility with existing x86 applications and lets them run at the processor's full speed. As a result, AMD positions the Opteron as the ideal solution for both 32-bit and 64-bit environments. In contrast, the Itanium 2 architecture has been optimized for 64-bit applications and doesn't provide native x86 support. The Itanium 2 runs 32-bit applications by translating their x86 instructions on the fly, resulting in 32-bit performance that falls far below that of the fastest Xeon processors.

Intel says that the 32-bit performance of the 1.5GHz, 6MB L3 cache Itanium 2 processor will improve to levels similar to that of the 1.5GHz Xeon processor with the release of Intel's IA-32 Execution Layer as part of Service Pack 1 (SP1) for Windows 2003, Enterprise Edition and Windows 2003, Datacenter Edition. The Execution Layer will perform the same instruction translation that the Itanium 2 currently executes internally. However, according to Intel, the Execution Layer does so much more efficiently. To date, weak 32-bit performance has hurt Itanium's sales because many applications aren't yet available in 64-bit versions, and buyers might be reluctant to port their proprietary applications to the new architecture right away. For this reason, Intel still recommends using Xeon-based servers for mission-critical 32-bit applications.