AMD's Opteron Dual-Core Processors

The Plan

The elegance of AMD's dual-core CPU lies in the method used to link the two processing cores on the chip. The plan for the 64-bit chips included a dual-ported System Request Queue and Crossbar, technology designed to directly connect one processor/cache core--and, eventually, two of them--to the integrated memory controller and HyperTransport pipelines. Now that dual-core technology is available, the second port can be put to use.

As with earlier Opterons, the 875 chip has a dedicated memory controller, so every new processor socket added to a system board increases the memory bandwidth by another 144-bit, 6.4-GBps data path. Although the 875's two cores share the same memory controller, the vendor says this impacts memory performance by less than 10 percent.

AMD's approach differs dramatically from Intel's Hyper-Threading technology, which only emulates a second virtual processor and lets two threads execute in parallel on a single core. Although this has proved to be a good use of processing cycles, both threads share the resources of a single processor. Intel also has introduced a dual-core Pentium for the desktop, but the design requires an entirely new support chipset and continues to rely on Intel's aging front-side bus technology.

Chip Tests

To get a feel for the dual-core Opteron, I went into our Green Bay, Wis., Real-World Labs® and built a test workstation from components supplied by AMD, including two Opteron 875s with coolers, eight 512-MB sticks of Kingston DDR 400 Registered ECC memory and a Tyan S2895 Thunder K8WE dual 940 motherboard. The Tyan Thunder is ideally suited to this application--it's based on a combination of NVidia's nForce Professional 2200/2050 PCI-Express chipsets and AMD's 8131 PCI-X HyperTransport Tunnel technology.

The motherboard came preconfigured and burned in, so I was spared the pain of struggling with the complex and marginally documented process of setting up a first-generation system. I added an NVidia 6800 PCI-X/x16 video card, EPS 12V/SSI-class power supply, DVD-R and a tower case capable of mounting and cooling a 12x13-inch workstation motherboard. For main storage, I used two Western Digital WD740GD 10,000-rpm, 74-GB Raptor SATA drives, and though the system offered on-board RAID support, I chose to use single drives to prevent contamination between the 32- and 64-bit versions of Windows to be tested.Watch It Run

The completed build started on the first try, so I loaded a copy of Windows XP Professional x64 Edition (RTM Build 1830, March 30, 2005) and installed the 64-bit device drivers. Not surprisingly, the NVidia chipset and integrated devices had excellent driver support; NVidia has been hip to 64-bit computing from the start and regularly updates its 64-bit video drivers for both Windows and Linux. For comparison, I did a similar build of 32-bit Windows XP Professional on a second drive and patched both systems up to current levels.

With the operating systems in place, I used the 32- and 64-bit versions of SiSoftware's Sandra Professional 2005 to confirm that the memory and processors were being properly recognized, then ran a series of five arithmetic benchmark cycles to determine average integer and floating-point performance. These tests were conducted using Sandra's default configuration, which enables multithreading and assigns only one thread per core.

Impressive Scores

In 64-bit Windows, our dual-core/dual-processor Opteron 875 system reached an impressive Dhrystone ALU (Arithmetic Logic Unit) score of 43,244 MIPS, a Whetstone FPU (Floating Point Unit) score of 15,847 MFLOPS and a Whetstone SSE2 (Single SMID Extensions 2) score of 18,320. These turned out to be only a few percentage points higher than the scores of 41,024 MIPS, 14,008 MFLOPS and 18,095 SSE2 MFLOPS I experienced running the exact same series under 32-bit Windows XP.

I executed the same 32-bit series on my existing dual-Opteron 244 workstation for a reality check and came up with scores of 16,102 MIPS, 5,699 MFLOPS and 4,764 SSE2 MFLOPS. These numbers show the 875 offered a 155 percent increase in 32-bit integer performance and a 146 percent increase in floating-point performance. The biggest difference between the old and new systems showed up in the SSE2-based floating-point calculations, where the 875 showed a performance increase of 280 percent over the older 244 processor and indicated some substantial generational improvements in SSE2 processing during the past 18 months.

AMD has benchmarks pitting the Opteron 275 against the more modern, single-core/dual-processor systems that show a 70 percent to 90 percent performance increase over the 252 and 248 chips. A quick series of tests I ran using Photoshop CS to manipulate 100- and 250-MB image files verified the increase in real-world performance I had been expecting from multiprocessor-savvy applications. With both environments configured identically for memory use, multithreaded filters like Radial Blur, Find Edges and Median Noise Reduction executed in less than half the time required by the dual-244 based workstation.

These numbers don't tell the whole story. In reality, only multithread-capable applications or systems requiring heavy multitasking will see the full benefits of dual-core processors. With this in mind, AMD is gearing its initial dual-core production toward server-class chips to provide product for server, blade and clustered applications where the issues of performance, power consumption and processor density play heavily. Tier 1 vendors Hewlett-Packard, IBM and Sun Microsystems have publicly validated AMD's dual-core philosophy, announcing plans to introduce a number of systems based on the new Opteron processors.

For now, dual-core systems are best suited for computationally intensive applications that require multiprocessing systems, such as large databases and busy Web servers. Memory-intensive tasks like mail and file serving, as well as the vast majority of single-threaded applications available to the business and consumer user, may still be best served by single-core systems, at least until software developers begin shipping multithreaded, 64-bit programs.Steven Hill owns and operates ToneCurve Technology, a digital imaging consulting company. Write to him at [email protected].

Enterprise users planning the jump should note that many software vendors consider dual-core devices to be the same as two physical processors. For example, IBM's licensing agreement for Informix Dynamic Server 9 has a clause that states "With multicore technology, each "core" is considered a processor. For example, in a dual-core chip, there are two processors; IBM software products, which are priced "per processor" will require two licenses when applied to a chip that uses dual-core technology." This position usually includes Intel's Hyper-Threaded, single-core processors, and appears to be standard for vendors that charge by the processor. So even though you may not save money on licensing you might be able to opt for less-expensive, four-way/dual-core servers over pricier eight-way/single-core models.

In a curious move, Intel has chosen a different marketing tack and is focusing the release of its first dual-core "Extreme" EMT64 Pentiums to the consumer channel--a market that enjoys very few applications that can actually take full advantage of dual-core processors. Even more interesting is Intel's dual-core road map, which places its next focus on the already-struggling Itanium line and doesn't address the server-class Xeon until some time in 2006. On the other hand, AMD isn't ignoring the consumer market potential for dual-core processors and has announced the upcoming release of a line of Athlon64 X2 processors in 4200+, 4400+, 4600+ and 4800+ versions; scheduled to begin shipping in June of 2005.