Quad-core Xeons use small power plant to maul Opterons
When Fully buffered DIMMs attack
A recent benchmark comparing quad-core chips asserts that Intel Xeon-based servers have an edge in performance but are complete power hogs when compared to the dainty energy appetite of AMD Opteron-based servers.
The test was conducted by the computer performance consulting firm Neal Nelson & Associates. It showed Intel Xeon-based servers reaching 14 per cent higher throughput performance over similarly configured AMD Opteron-based servers, which consumed 41 per cent less power.
Interesting results, considering Intel's bravado about power efficiency with the Xeon chips. According to the research firm, Intel's claim isn't necessarily incorrect from a narrow perspective. It's the wide angle that bites Chipzilla.
"By themselves the Intel processor chips may use less power, but all current Intel Xeon servers require the use of Fully Buffered memory modules," said Neal Nelson, the test's designer. "These FB-memory modules appear to consume more power than the DDR-II memory modules used by the AMD-based servers. The result is that in many cases an Opteron based server uses less total power than a Xeon-based server."
Those of you following the Xeon vs. Opteron debate will recognize this as a familiar point.
The benchmark was performed using 32 computers to submit 500 individual transactions to a server running Novell's SUSE Linux, Apache2 and MySQL. The firm said the test has a complex multi-user load with a large memory footprint, a high volume of context switches, significant network traffic and substantial amounts of physical disk I/O. Servers put through the wringer used 1GB memory modules in 4, 8 and 16GB memory configurations.
When database working sets were small enough to fit into the servers' kernel disk buffer cache — using virtually no physical disk I/O — Xeon servers had up to 14 per cent greater throughput than Opteron. When the size of the database was too large to fit in the kernel disk buffer cache so substantial physical disk I/O came into play, the Xeon processor advantage dwindled to 3 per cent greater throughput.
When tasked with identical levels of transactions, Opteron-based servers consumed up to 32 per cent less power than Xeon servers. When the systems were idle, Opteron servers consumed up to 41 per cent less power than Xeon servers. Nelson points out that database servers are idle up to 80 per cent of the time.
The results were similar to the firm's benchmarks of dual-core Opteron- and Xeon-based servers last August.
In his personal view, Nelson said the real crux of the results is Intel's use of energy hungry FB-DIMMs.
"Intel shot itself in the foot here. And there's really no indication of using a different bridge for multi-socket servers."