Since the release of Woodcrest earlier this year, Intel has been resurgent in the dual processor server market. In earlier articles we explored both the performance and power characteristics of Woodcrest, and in each case it was a substantial improvement over the Dempsey microprocessor. More to the point, Woodcrest is clearly a very potent competitor and has helped Intel retake marketshare lost to AMD.
As part of their strategy to regain momentum in the server market, Intel has decided to press home their advantage by aggressively moving to a four core solution as soon as possible. Clovertown is the first x86 four core MPU and a socket compatible upgrade for platforms based on the Blackford chipset, including both Woodcrest and Dempsey. Rather than designing a four core microprocessor implemented in a single die, Intel opted to package together a pair of processors to create Clovertown. The changes required were largely outside of the die of the microprocessor, which makes a new version both quick to release and relatively inexpensive. The end result is the third in a series of substantial performance improvements for server buyers in 2006.
The one draw back of increasing the number of cores in the same platform is that it decreases the per core bandwidth. Clovertown actually decreases the absolute available bandwidth, due to additional cache coherency traffic, so a 2x increase in performance is not expected on any realistic workload. In reality, the performance gains will vary depending on the relative memory bandwidth and computational requirements of a given application. The general expectation set by Intel is that Clovertown will improve performance by anywhere from 20-75% across a wide range of workloads, with the vast majority falling in the 25-50% range.
Clovertown will initially be released in four bins: 2.66 and 2.33GHz with a 1.33GT/s front side bus and two lower end models at 1.86 and 1.6GHz with a 1.06GT/s bus. The highest bin has a 120W TDP, while all other models are at 80W. The 2.33GHz device will be priced at parity with the existing 3GHz Woodcrest, while the 2.66GHz part will be marketed as a higher bin and priced accordingly.
For this review, we will be using a single “Starlake” system, based on the Blackford chipset, with two different pairs of processors: Woodcrest 3GHz and Clovertown 2.33GHz. Note that while these two processors are priced at the same level, a higher performance Clovertown exists and will be used for most benchmarks, but does have higher power consumption. The system used is described below in Chart 1, additionally, hardware prefetch was enabled for all benchmarks. Both power and performance numbers will be collected and analyzed for four Windows x64 based benchmarks: POV-Ray, Sungard, XML Test 1.1 and SPECjbb2005. Performance is measured and averaged over three runs of each benchmark, and power is measured using an Extech power meter, which provides ‘watts at the wall’.
Chart 1 – System Configuration
Discuss (One comment)