The processor comparisons indicated that while the Xeon quad-core-based servers delivered up to 14% higher throughput than did AMD's quad-core Opteron-based models, Opteron servers were as much as 41% more power efficient. While the study may reflect questionable methodology, it also begs questions about which kind of memory best serve processors.
Apples to oranges benchmark
Indeed, Neal Nelson's results are problematic. The tests compared AMD's newest technology, the Opteron 2350 2.0 GHz quad-core processors released in Q4 2007, with Intel's older 65-nanometer (nm) Clovertown E5335/E5345 processors, which became available in Q4 2006.
Nelson could have tested newer Intel quad-core processors, such as the X5365 and L5335 released in August 2007, which had enhanced energy efficiency and performance, or the further enhanced 45-nm Intel 5400 quad-core processors released in November 2007.
Nelson said he used the older Intel quad-core processors because when he requested newer parts for testing from Intel, the company declined. "The 65-nm parts are being shipped in volume to customers today, so those are the parts that I was able to borrow to run these tests."
But while AMD's Opteron Barcelona processor officially launched this September, the chip hasn't been shipped, and Nelson declined to say where he procured it.
AMD decided not to go forward with widespread shipments of Barcelona after it discovered an errata, explained John Fruehe, AMD's worldwide market development manager of server and workstation products. "We don't believe customers would even see it, but we are conservative and would rather be safe than sorry, so we held back on bulk shipments," Fruehe said, adding that AMD is on track for volume shipment in April with revision B3.
Despite the use of older Intel technology, Nelson maintains he is vendor neutral and was not remunerated by AMD. In fact, he gave Intel some esteem, saying, "Intel will almost always report times that are a lot better than AMD when the primary task load is bigger than 2 MB [the size of the AMD cache] and smaller than 4 MB [the size of the Intel cache], and when there are few interrupts -- low network I/O -- and no disk I/O."
An Intel spokesperson declined comment on the test results, but cited the more favorable Standard Performance Evaluation Corp. (SPEC) benchmarks.
Memory efficiency at issue
Despite the questionable processor comparison, the type of memory Intel uses, Fully Buffered DIMM (FB-DIMM), is known by memory manufacturers such as Micron Technology Inc. to consume more power than the DDR2 DIMM memory used by AMD. The distinction matters because it can counteract the power efficiency built into a CPU.
"By themselves, the Intel processor chips may use less power, but all current Intel Xeon servers use of FB-DIMM memory modules. These modules appear to consume more power than the DDR-II memory modules in AMD-based servers. The result is that in many cases an Opteron based server actually uses less total power than a Xeon-based server," Nelson said.
On average, the CPU accounts for 35% of server power consumption, and memory sucks up 15%, according to research b Midvale, Utah-based Burton Group.
Nelson said that even if he had used the newest-model Intel processors, the power efficiency would show only about a 2% improvement, because the latest Intel processors still rely on FB-DIMM memory.
"For big and complex workloads, like Apache and Oracle, with lots of network activity and some disk I/O, Intel may be somewhat faster than AMD, but because Intel insists on FB-DIMMs, AMD will be much more energy efficient," Nelson said.
"For workloads that are bottlenecked on disk I/O, both machines will deliver almost identical throughput, and the AMD machine will be more energy efficient," Nelson said. "I saw this same pattern for the current [core architecture] Xeons in both dual- and quad-core configurations."
Further demonstrating the effect of memory , Sun Microsystems Inc. has compared its Sun Fire X4150 powered by the quad-core Intel Xeon processor 5300 series with an AMD Opteron.
Sun's comparison indicated that in Intel Xeon 5300 series-based servers, the CPU consumes less power but has longer memory latency and an FB-DIMM consumes 10.4 watts of power per module. In contrast, the AMD Opteron CPU consumes more power but offers shorter memory latency, while the DDR2 DIMM consumes only 4.4 watts per module.
Intel has said that FB-DIMMs may ultimately provide lower total cost of ownership than their alternatives because of greater reliability and memory capacity.
To reduce memory power consumption, Micron introduced the energy-efficient memory module in 2007 Aspen Memory, which can reduce server power consumption by 60%, the company claims.
"A server farm with 1,000 blades is concerned about power beyond anything else. These folks could save tens of millions of dollars buy using a memory technology we offer, and not at a premium," said Brett Williams, Micron's senior segment marketing manager of DRAM. "You don't have to use the power you use now."
Micron devised a power calculator that allows data centers to calculate how much power their memory consumes.
Let us know what you think about the story; email Bridget Botelho, News Writer.
Also, check out our news blog at serverspecs.blogs.techtarget.com.