Hyper-Threading Scalability Remains to be Proven
- By Stephen Swoyer
- March 19, 2002
As is common with launches of new processor lines, the new Intel Xeon MPs do not deliver on their performance promises yet, even in benchmark situations.
Intel says its new generation of SMP processors gives up to a 30 percent improvement boost, but the company also says that the new Hyper-Threading capability of the Xeon MPs can add up to an additional 30 percent performance improvement if an application is coded properly.
Intel's best public benchmark at the Xeon MP launch last week boasted a 36 percent performance increase over a similarly configured Pentium III Xeon system running an SAP application. (A benchmark showing a 40 percent improvement used more memory than the Pentium III Xeon system it was compared against.)
Intel spokesman Bill Kirkus says customers' "mileage will vary greatly, but we're saying up to a 30 percent boost on Xeon MP. That comes from two areas, the integrated on-die L3 cache and the faster clock speeds, because we’ve gone from 900 MHz to 1.4, 1.5 and 1.6 GHz.”
Hyper-Threading, which allows one processor to behave as two logical processors by simultaneously handling multiple application threads, isn't adding much additional performance yet -- by the standards of Intel's benchmarks thus far. Assuming Intel gained 30 percent from the clockspeed and L3 cache on the mySAP Supply Chain benchmark, that leaves a 6 percent boost from Hyper-Threading.
“We’ve been trying to stay pretty conservative here, especially in the data center where everything is about money. Even a one percent performance benefit in the data center is great for people,” Kirkus says. “Other benchmarks will soon be published. One lab just did one specific test on a two-way and they saw 40 percent. Your mileage may vary, so it will depend on the application and how it’s tuned.”
Kirkus contends that a 30 percent performance boost from Hyper-Threading alone remains possible on applications that are multi-threaded to a high degree.
“A lot of the enterprise class applications that run on two-, four-, eight-, or 16-way systems already to a degree have multithreading capability because they run on multiple processors. So when you add Hyper-Threading, you really get performance because those apps are already multi-threaded,” Kirkus says.
Analyst Nathan Brookwood of microprocessor consultancy Insight64 says the Xeon MP performance picture should become clearer once Intel releases its next version –- codenamed “Gallatin” –- later this year. Gallatin, which is based on a .13 micron process, will be able to reach higher clockspeeds than Xeon MP.
According to Brookwood, this is important because Xeon MP’s Netburst architecture, which offers memory and I/O bandwidths that far outstrip those of the aging Pentium III Xeon bus, also requires more instructions per clock cycle. The Xeon MP must run at faster clockspeeds relative to the Pentium III Xeon to achieve the same processing power.
“The Netburst micro-architecture takes a hit in terms of instructions per clock. Typically, in the desktop space, you need 30 percent more clock rate just to stay even with the desktop Pentium III,” he explains, granting that “processor clock rates are only a small part of the overall equation … in the server space.”
Stephen Swoyer is a Nashville, TN-based freelance journalist who writes about technology.