Core Microarchitecture Performance: Woodcrest Preview

Pages: 1 2 3 4 5 6 7 8 9 10

Conclusion

Up until today, with the release of the first Bensley platform, the server market was really a one horse race in terms of performance and power efficiency. Paxville, Intel’s first dual core product, was a quick, dirty and ugly implementation that used up lots of power and gave little back in the way of performance. The combination of Dempsey and Blackford will significantly improve performance and should put Intel back on the map for servers. Dempsey may not beat the K8’s performance, but it should be able to get within 10-20% on most benchmarks. Unfortunately, Dempsey does little to reduce power consumption. However, Intel’s goal is not to achieve parity with AMD’s existing products. At the end of the day, everyone wants to know whether Intel’s vaunted Core microarchitecture is up to the task of reclaiming the server market from AMD. Can Woodcrest really live up to the hype?

Intel gave us some performance estimates for Woodcrest that seemed to answer in the affirmative. But what is the story in the real world?


Figure 9 – Real World Relative Performance

Intel’s performance estimates show an 11-50% increase in performance from Dempsey to Woodcrest. These numbers are a bit higher than ours, but they are not unreasonable. We measured 15-40% speed ups across a wide range of Windows applications, including ray tracing, Monte Carlo risk analysis, XML processing and Java based mid-tier OLTP workloads. While our numbers are lower than Intel’s and our Dempsey system runs at a slightly lower clockspeed (3.4GHz versus 3.73GHz), there is still a lot of performance left on the table for Woodcrest. Current compilers, JVMs and applications have not been optimized for Woodcrest, and our system is using slower FBD-533 memory, rather than FBD-667. Furthermore, our performance is somewhat lower because we have 1 FB-DIMM per channel; two FB-DIMMs are needed for the best performance.

Given all this information, Woodcrest should outperform competing F-stepping Opterons (which should hit 2.8GHz, or maybe even 3.0GHz) by around 10-30% in most applications, when it arrives in June. The icing on the cake is a subject we have not even discussed: power and energy. Our anecdotal observation is that Woodcrest systems use half the power of Dempsey systems, while providing much better performance, but we will have some real data in the near future. The bottom line is that Intel is well positioned to regain control of the dual socket server market from June, till the start of 2007.

Acknowledgements

I would like to thank the following people for all their help:

  • Scott McLaughlin
  • Waseem Ahmad
  • Trevor Lawless
  • Mike Houlihan
  • Larry D. Gray – I appreciate the hospitality
  • Jerry Baugh
  • Sanjay
  • Christian Anderka – Thanks for the credit card
  • George Alfs
  • Kim Lichong
  • Microsoft – For providing us with the software needed for this review
  • My Fellow Reviewers
  • All of the presenters
  • and anyone else I may have forgotten, or purposefully omitted so they don’t get into trouble

Pages: « Prev  1 2 3 4 5 6 7 8 9 10  

Discuss (22 comments)