AMD Phenom Processor Update – Memory Bandwidth

AMD Phenom Processors Don’t Have Crappy Memory Bandwidth Afterall

Not much has been going right for AMD in recent weeks, but I’m happy to say one of the most popular benchmarks used in testing processors wasn’t ready for Phenom when the media hacked apart the poor processor apart on launch day. It seems SiSoftware Sandra XII was not correctly optimized for the Phenom processors! We contacted SiSoftware after we were informed by AMD that nearly every memory bandwidth score on the internet was incorrect. Why AMD waited weeks to let us know this is beyond me, but nothing is shocking in this industry.

SiSoftware has this to say about what happened:

“If you check the benchmark breakdown you see it is now using one thread per core instead of one thread per package (physical CPU) like earlier versions did. If you disable multi-threading you will see the same score. I would call it an optimization; some CPUs get better performance with one thread per core (Yonah, K8 DDR2, Phenom, future Intel CPUs � while others with one thread per package (Pentium D, Merom, Penryn, etc.). It is not always obvious from the datasheet which code-path is better without a test system, and there were no Phenom/Barcelona developer boxes before we released original XII. We try our best to get information and support future devices before release, sometimes the released device differs from preliminary datasheets, etc…” – Sisoftware 12/19/2007 to Legit Reviews

Armed with Sandra XII and Sandra XII SP1, I went back and re-benchmarked our test system which happens to be running a BIOS with the TLB patch enabled in it.

Phenom 9900 on Sandra XII with the TLB Patch:

With Sandra XII our memory bandwidth was found to be 4789 MB/s on the Integer and 4805 MB/s on the Floating point.

Phenom 9900 on Sandra XII SP1 with the TLB Patch:

On Sandra XII SP1 our memory bandwidth was 9624 MB/s on the Integer and 9637 MB/s on the Floating point, which is double what Sandra XII was showing. It is obvious from the results that using one thread per core instead of one thread per package is the best way to benchmark Phenom. Sandra XII SP1 seems to fair for Phenom processors and has the performance back to where it should be. The reason I’m bringing this up is because our article on the Phenom processor contains data that is inaccurate and it needs to be updated. This also impacts other major hardware review sites and I’m not going to point any fingers and say who needs to update their charts, but dozens of sites need to. When it comes to Legit Reviews, I have gone back and made the changes to our original Phenom article as there is a significant difference in the results. LR strives to keep our benchmark numbers correct and since this was a big change it needed to be explained!