Intel's Nehalem Performance vs. IBM’s POWER Architecuture

My previous post “Intel’s Nehalem Processor – Performance Evolution or Revolution?” looked at Intel’s new generation processor architecture, codenamed Nehalem, and its performance in relation to other Intel processors, as there were indications that Intel had made more than the usual improvements in performance over previous generations. And the analysis did back this up.

Given that, the question then arises, how does the Xeon 5500 Series iteration of the Nehalem architecture stack up in a wider sense on other workloads, and in particular against what many see as the benchmark when it comes to per core performance, the IBM POWER processor series?

This post will look at results from the TPC-C Benchmark, which is a favorite of IBM’s and where many believe the POWER architecture performs best relative to the competition, to see if we can answer that question.

Chart 3 below, is similar to the charts in the previous post, in that it compares the average performance for a given processor type at a given configuration core count. There may be more than one actual result for each bar, but for simplicity, the average of all results is taken and plotted. In addition, the averages are further categorized into the number of cores tested. So a processor bar may appear more than once in the chart, representing the average performance for all configurations using that number of processor cores.

The chart features current TPC-C results since the beginning of 2007 for configurations using the Intel Xeon and IBM POWER processors, in order to get a meaningful sample to examine.

Three things are immediately apparent.

• Firstly, the per core processor performance differential between Intel’s Nehalem (green) and previous generation Intel server processors (grey) is similar to the TPC-E Benchmark, which we examined in the previous post.
• Secondly, the average per core performance of the Intel processor is getting into the territory of the IBM POWER processors (blue).
• Finally, IBM is showcasing performance with significantly higher core counts to that of the Intel processor, which is an area of differentiation that IBM points to, between the two offerings. Although, when the ‘Nehalem EX’ processors for servers with four or more sockets are introduced we can expect to see Intel challenging IBM in the 16-32 core space and beyond.

At this point it is worth noting that IBM’s results are relatively old, with the most recent test being over a year ago and no current generation POWER6+ processor is featured. Also, there are some critics in the industry that feel the TPC-C benchmark itself is old and other benchmarks are more relevant.
 
So, the next post will examine results for the SAP 2-tier benchmark, to see what a change in workload reveals.

Comprehensive consolidated industry standard performance data is available from Ideas International via two sources.

The free Benchmark Gateway service, which features the Top Ten results for major industry standard benchmarks and the subscription service Competitive Profiles that provides comprehensive data coverage of the features, pricing, and performance characteristics of a wide range of enterprise IT products and services.

Gary Burgess
Senior Vice President, Research & Operations
 Follow Gary on Twitter

Comments

[Thibaud]

Gary,

price/perf in tpc benchmark is mostly driven by the cost of the storage, not the cost of the server.

Therefore, it is not representative of the value of the server, specially in virtualized environements on large servers.

Regards,

Thibaud

[Thibaud]

Gary,

be aware that price per transaction as calculated by tpc takes into account the whole solution : server, storage, software licences. In tpc, the largest contributor to cost is storage, by far. Check the details of the benchmark.
Therefore, it is not representative of the cost of the server, because so many disks are required to generate the workload. For example, for HP Nehalem server DL370G6 tpc benchmark, storage price is ten times higher than server.
For a given customer, actual TCO and ROI may be completely different than tpc price/perf, specially in virtualized environments on high-end servers. Best plateform may not be what was expected.

[Gary Burgess]

Antonio,

The TPC-C analysis looked at POWER6 results only. There were no POWER6+ results because IBM and/or Bull have not yet released any TPC-C results using the POWER6+ processor.

My next post intends to look at the SAP 2-tier benchmark, where some POWER6+ results are available.

Gary

[Gary Burgess]

Hans,

Thanks for your post. Yes you are correct, the TPC benchmarks do have a price perspective, which other benchmarks don't often have. My focus of this series of posts is on performance and why I didn't consider the pricing angle.

However, you make a valid point and so for the record the pricing details of the POWER6 and Nehalem processor results are outlined below.

4-core POWER6 results consisted of the following:

* Bull Escala PL1660R with 404,462.5 tpmC at a price per transaction of $3.51/tpmC
* IBM Power 550 Express with 276,383 tpmC and $2.22/tpmC
* IBM Power 570 with 404,462.5 tpmC and $3.50/tpmC

4-core Xeon E5520 result was a:

* HP ProLiant ML350 G6 with 232,002 tpmC and $0.54/tpmC

8-core POWER6 results were:

* Bull Escala PL860R 629,159.2 tpmC for $2.49/tpmC
* IBM Power 550 Express 629,159.2 tpmC for $2.49/tpmC

8-core Xeon X5570 result consisted of

* HP ProLiant DL370 G6 with 631,766 tpmC for $1.08/tpmC

16-core POWER6 results details are:

* Bull Escala PL1660R with 1,616,162 tpmC for $3.54/tpmC
* IBM Power 570 with 1,616,162 tpmC for $3.54/tpmC

64-core results consisted of:

* Bull Escala PL6460R with 6,085,166tpmC for $2.81/tpmC
* IBM Power 595 with 6,085,166tpmC for $2.81/tpmC

Based on these outcomes, the Nehalem processor configurations have a lower price per transaction (half or less) to that of the IBM POWER6 configurations.

Gary

[Antonio Carlos Navarro]

Is this study comparing IBM Processors Power6+ or only the Power6 generation? Thank you!

[Hans Jahn]

Gary,
TPC-C is a price/performance benchmark which means that a mix of best performance and cheapest price is the target while non TPC benchmarks like SAP-SD, VMark and others are focused on performance and price agnostic.

Comparing different OS's and DB's as well as the price/TPC-C including discount strategies are not reflected in your comparison.
A reader might think that you compare green apples with green apples.

Regards Hans