AMD Pushes Opteron Clocks to 3 GHz, Boasts of Benchmarks

Published: April 4, 2007

by Timothy Prickett Morgan

With Intel coming on pretty strong with its "Clovertown" Xeon 5300 processors and Advanced Micro Devices having to get to June or July before it gets its own quad-core "Barcelona" Opteron chips out the door, the company has to do something to make its Opterons appealing. To that end, AMD has pushed up the clock speeds on the dual-core Rev F Opterons to 3 GHz to compete better in the dual-core X64 chip space.

According to Pat Patla, director of Opteron marketing at AMD, the new 3 GHz chip is a so-called Special Edition, or SE, part. This means that it has a thermal design point of 120 watts, not the 95 watts of the standard Rev F Opteron parts. As part of the rejiggering of the Opteron lineup, the 2.8 GHz SE part will be delivered in a standard 95 watt TDP, and will presumably see a price cut. AMD will also presumably charge a premium for the faster 3 GHz part as well. While AMD has been shipping the faster chip to partners for a few weeks, it is not yet available through the retail and OEM channel and pricing has not been set for it yet.

The faster Opteron chip is available for the two-socket servers as the Opteron 2222 SE and in machines with four sockets or more as the Opteron 8222 SE. Patla says that AMD will eventually deliver the faster chip in the Opteron 1000 series for single-socket servers (probably in about a month), and says further that the company will crank the clock one more time on the dual-core Rev F Opterons to 3.2 GHz--again, as an SE part with the higher 120 watt TDP and pushing the 3 GHz part down to a regular Opteron part and a 95 watt TDP.

After that, AMD will be focusing on the quad-core Barcelona chips and how they stack up to the Intel Clovertowns, which are really two dual-core "Woodcrest" Xeon 5100 chips that are crammed into a single chip package and which share a single socket and front side bus.

In the meantime, AMD is keen on proving that its has the best performance in the dual-core X64 chip space, and to that end it took out a full page ad in the Wall Street Journal that sniped: "Saying you're the world's best processor is one thing. Actually being the world's best processor is another." Patla says that there has been "a lot of confusion" about whether Intel or AMD has the fastest processor in terms of measured performance, and the advert was an invitation to review the SPEC CPU performance tests that Intel and AMD have done to prove the mettle of their chips.

"Granted, AMD does not have the 40 percent to 50 percent advantage that we used to hold, but we nonetheless have the best performance," says Patla.

On the SPECint2006 test, the new dual-core 3 GHz Opteron 2222 SE chip has a rating at 14.9 SPECint2006 Base running the test on Novell's SUSE Linux Enterprise Server 9 and QLogic's PathScale 3.0 compiler suite; this test was run on a whitebox server with a Tyan S4985 motherboard. A Bull NovaScale machine with two (not one) dual-core Woodcrest Xeon 5160 chips with a 3 GHz clock speed running Windows Server 2003 X64 Edition and Intel's own C++ Compiler 9.1 is rated at 16.6 on the same base test. The Intel setup has twice as many processors and only does 11 percent more work. A similar box from Apple Computer had the same 16.6 SPECintBase rating running on Mac OS X 10.4.8 and using the Intel compilers. A Dell Precision 690 workstation using a single Xeon 5160 chip had a rating of 17.5 on the SPECintBase test; this machine used the Intel compilers on the 64-bit version of Windows XP Professional. This is an awful big swing in terms of performance, and the Dell numbers stick out like a sore thumb compared to other machines tested. A Fujitsu Celsius R540 workstation with two Xeon 5355 Clovertown processors running at 2.66 GHz was rated at 16 on the test. Intel is putting four times as many cores as the new Opteron 2222 SE in the box, but is only delivering 7 percent more integer processing power.

On the SPECfp2006 floating point processing test, the same dual-core Opteron 2222 SE box (only one socket) had a base rating of 14.3, which was nearly identical to an Apple Xserve with two 3 GHz Woodcrest processors (that's four cores), which came in at 14.4 on the test. And machines with the quasi quad-core Clovertowns, which run at a slower clock speed, only did a little bit better on the test. A Fujitsu Celsius R540 workstation with two Xeon 5355 processors running at 2.66 GHz was rated at 16.1 on the SPECfp2006 base test. That's four times as many cores as the Opteron 2222 SE box for a 13 percent performance boost.

On the SPECfpRate test, which measures throughput as well as number-crunching power, the Clovertowns have a bit more of an advantage compared to the Rev F Opterons. The Clovertown machines are rated in the high 50s and pushing into the 60s on the SPECfpRate test, while the Rev Fs are pushing into the 50s. But the Intel setups have twice as many cores.

If you adjust for per-core performance, AMD is coming out ahead. Which was the point AMD was trying to make, but which it did not convey very well in its ad. Still, for some workloads, Intel is putting more performance into a single socket. But it is perhaps not as much as we might have been led to believe. Moreover, the differences in the performance between the Intel and AMD chip families may say more about the SPEC CPU workloads and how they respond to threads than they do about the innate raw performance that real end users can expect to see on real workloads. As was the case with ever-increasing megahertz and gigahertz in the 1990s and early 2000s, performance did not always scale with clock speed. And there is no reason to believe that performance will always scale with core count times clock speed.

                     Post this story to del.icio.us
               Post this story to Digg
    Post this story to Slashdot