A server configured with the new AMD low power Opteron “Shanghai HE- processors used 13 to 21 percent less power while delivering better throughput when compared to a virtually identical server configured with Intel Xeon “low power- processors. The test results were collected during extensive testing by Neal Nelson, an independent computer performance consultant.
“The AMD based server beat the Intel based server in our tests. We processed millions of transactions against a real database and we measured the total throughput and the total power used by the servers. We are certain of the accuracy of our test results and we provide our customers with a money back guarantee. If a system under-performs in the real world, we will take it back and refund the purchase price,- commented Neal Nelson, president of the consulting firm.
The servers were tested with three different transaction workloads: a calculation intensive workload, a disk I/O intensive workload and the “idl”state when the server was powered up and waiting for transactions to arrive. The entire test suite was run twice, first with 4 gigabytes and then with 16 gigabytes of main memory.
For the calculation intensive workload the Opteron reported 1.0 to 6.3 percent higher throughput. The disk intensive throughput tests ranged from 1.1 percent favoring the Xeon to 5.0 percent favoring the Opteron. The power consumption, however, favored the Opteron in every test with a range of 13.1 to 20.8 percent. When the power readings were normalized by throughput the Opteron advantage ranged from 21.8 to 26.8 percent.
The servers used in this test had virtually identical configurations. The only notable differences were that the Intel server had an 8 percent faster clock frequency with 2.5 versus 2.31 gigahertz, and the Intel server used “Fully Buffered- memory modules (FB-DIMMs).
Nelson has observed that for several years Intel has been asserting that it’s central processing units (CPUs) are more power efficient than comparable CPUs from AMD. Intel, however, does not discuss the fact that, for the current generation of Xeon server CPUs, the Intel architecture requires an external memory controller chip as part of the server chipset (this is not required for the AMD CPUs) and that Intel mandates the use of FB-DIMM-based memory technology which draws more power than the DDR2 memory technology used by AMD processor-based servers.
The net effect of these architectural differences is that an Intel based server can draw more power at the platform-level when compared to an AMD based server that delivers comparable throughput. A server’s total throughput and total power consumption are much more important than whether some individual computer chip consumes more or less power.
The Nelson test results should not be confused with power usage test results from the Standard Performance Evaluation Corporation (SPEC). The SPECPower test was created by a committee of engineers from computer manufacturers (including Intel). SPEC issues a disclaimer which specifically states that the test results may not predict real world experiences.
A free white paper with more information about these test results is available from Neal Nelson by email at neal@nna.com
