Intel Quad-core: Clovertown Performance Reviewed
Posted on: 11/11/2006 06:00 AM

Performance-per-Watt

Overall power usage is nice to know, but everyone is focusing their attention on actual "performance-per-watt" (PPW). Overall power usage is easy to determine, but unfortunately it is pretty difficult to quantify PPW on a platform or system level. There are many factors that go into determining PPW and results will vary depending on workloads and utilization.

SPEC is currently developing a new benchmark that will generate an overall PPW "score". That will be a decent metric for comparison, but to effectively quantify PPW, you will have to test with your own (or very similar) application, and do it at your estimated usage levels. In the end, who really cares what kind of power a machine draws at 100% CPU load running SQL Server if your machine will be running 40% loads with a Java application? Let me show you what I mean...

For this test I wrote a batchfile that runs our custom Black & Scholes kernel. I used the same number of steps (1,000,000,000) as the normal benchmark runs, but I changed the number of threads to simulate 25, 50, 75 and 100% CPU loads (1, 2, 3 and 4 threads for Opteron and Woodcrest, 2, 4, 6 and 8 threads for Clovertown and Dempsey), hence the stair-stepped lines on the graph. If we analyze the power usage and time to completion for each stepped load, we can make some conclusions about PPW for our test machines.



The Woodcrest Xeons turn in the slowest times in our Black & Scholes kernel benchmark at all load levels. The first thought would be that Woodcrest offers poor PPW, but if you look at the amount of power that you would save (at all load levels) over Opteron and Dempsey, you begin to wonder if it might be worth giving-up a little speed to save electricity ($$) in the long run. Obviously Clovertown offers the best PPW, in this test.

Therein lies the rough. As we've seen in this article, the Black & Scholes kernel test contradicts the performance results of every other test I've run. Basing PPW conclusions solely on this test would be a mistake. You may ask why I even bothered to include it, and that reason is to hopefully start a discussion on what PPW should actually show, what we should base it on (workloads, load levels, etc). Should I do similar power tests for every workload (I actually thought about doing it this way)? Or should we just wait for a standards body to release an "industry standard" test? I don't have an answer, but I think we're moving in the right direction, and I think there is a lot of discussion that needs to take place.


Printed from 2CPU.com (http://www.2cpu.com/contentteller.php?ct=articles&action=pages&page=intel_quad_core_clovertown_performance_reviewed,14.html)