X
Tech

Intel turns up the efficiency and performance heat on AMD

Intel's latest manufacturing refinements to the 65nm Core 2 architecture has produced server CPUs that drop power consumption, boost performance, and include virtualization enhancements.  These new refinements called "Stepping G" have also given higher performance and efficiency to Intel's desktop line of Core 2 processors allowing them to overclock past 4 GHz.
Written by George Ou, Contributor

Intel's latest manufacturing refinements to the 65nm Core 2 architecture has produced server CPUs that drop power consumption, boost performance, and include virtualization enhancements.  These new refinements called "Stepping G" have also given higher performance and efficiency to Intel's desktop line of Core 2 processors allowing them to overclock past 4 GHz.  On the server side Stepping G has given us the new Intel XEON X5365 and L5335 microprocessor.

The Xeon X5365 "Clovertown" quad-core CPU at 3.0 GHz with a TDP (Thermal Design Power) of 120 watts sets a new SPECint_rate_base2006 world record for two-socket servers with a score of 98.9 (peak 107).  According to TechReport's Scott Wasson, the new stepping slashes idle power consumption of the X5365 in half from 50W idle to 25W idle power consumption which is only 1W higher than the idle power consumption of the new low voltage L5335.

Note that Apple released a 3.0 GHz "Clovertown" back in April for their 8-core workstation but that chip had a TDP of 150 watts which is too high for normal two-socket servers.

The Xeon L5335 "Clovertown" quad-core CPU at 2.0 GHz low voltage processor has a TDP rating of 50W but no SPECint_rate2006 numbers are available at this time.  We can get an idea of how well it performs by looking at the older E5335 80W part which has a SPECint_rate_base2006 score of 78.5 (peak 85) if all the other components in the server match.  The 50W Intel Xeon L5355 will make it very hard for AMD to claim the performance/watt lead on their yet-to-be-released 95W Barcelona 2.0 GHz CPU.  Even if we take in to account the extra 29 watts on Intel's chipset because of the "North Bridge" memory controller on the motherboard and the 48 watt penalty of eight fully buffered memory DIMMs, it still gives Intel a 13W advantage in a two-socket server.

While no SPECint_rate2006 scores are published for AMD's Barcelona CPU yet, many industry analysts are expecting an AMD "Barcelona" 2.0 quad-core to lose to an Intel Clovertown 2.0 quad-core CPU on integer performance.  Considering the fact that AMD's Barcelona 2.0 GHz CPU is a 95W part that will likely have lower integer performance than an Intel Clovertown 2.0 GHz 50W part, it's understandable why AMD has declined to announce any integer performance numbers yet for their 2 GHz Barcelona.  AMD has claimed that they can get a SPECfp_rate2006 peak score of 69.5 for floating point performance on the 2 GHz Barcelona which is very impressive but that only addresses a niche HPC (High Performance Computing) market.

AMD will eventually launch a low-voltage 68W variant of the Barcelona 1.9 GHz quad-core CPU but the initial version of the 1.9 which most likely launches in September will be 95W.  Even then it will only be a narrowly defined victory in a two-socket CPU 8-DIMM configuration for the absolute lowest power consumption and it may or may not be compelling on performance/watt for typical server workloads.  AMD can't delay too much longer because they will soon be facing a much bigger monster in the form of Intel's 45nm Penryn.

From an HPC perspective the AMD Barcelona 2.0 GHz parts will be an absolute steal because it won't be able to command high margins due to its mediocre mainstream server computing performance.  That might be good news for scientists and researchers who need cheap and powerful HPC workstations but the modest margins won't fix AMD's financial woes unless they can significantly ramp up the clock speeds on Barcelona quickly.  For now it seems that all we can do is wait on news for AMD while Intel seems to be tick-tocking away on execution.

Editorial standards