New Patches Boost Bulldozer Performance in Windows 7

In October, AMD released various processors based on the Bulldozer core. Much hyped as the next best thing, these processors were shown to perform much worse than expected, partly due to the new multi-module architecture, and poor scheduling (ie no support) in Windows 7 and Windows Server 2008 R2. Microsoft have now released a pair of hotfixes which aim to improve support and hence improve the performance.

The problem is that Bulldozer processors are split into modules, and each module contains two integer cores and one floating point core - a set-up that Windows 7 was not optimized for. A processor that is generally described as octo-core actually has eight integer cores, but Bulldozer only has four floating point cores to back them up. The Bulldozer core is used for the latest FX processor, as well as Opteron 6200 server processors, so performance gains and losses affect a wide potential user base. Fortunately, Microsoft have now released two hotfixes, which need to be installed in a specific order and will then make better use of the Bulldozer architecture.

The first update to install is KB2645594. This addresses scheduling in lightly threaded applications, with the first 4 threads in an app being allocated to separate modules. This is in contrast to ordering the threads by the order of the integer cores, which could leave some modules fully loaded with others idle. In this case, spreading threads across as many modules as possible is more efficient than than using all the cores on one module and none on another.

Next, we have KB2646060, which addresses some power issues that also affect performance. If the idle portions of a partially used core go into idle state, the rest of that module is also affected. A bulldozer module is less likely to go into the C6 power state when this update is applied, which can potentially increase power consumption but will also increase performance.

VR-Zone have carried out some "before and after" tests for these patches, using a dual-Opteron 6274 system (8 modules, 16 cores per processor). The SIMD/AVX performance (tested using SiSoftware Sandra 2012) shows a performance drop of approx 0.9% after application of the patches. Inter-core bandwidth increased by 14%. The remaining benchmarks showed an improvement of between 0.6% and 4.9%. Discounting the inter-core bandwidth, which is disproportionate to the other results and is unlikely to be noticed in most real-world scenarios, the average performance boost in the remaining tests is approximately 1.85%. This might not sound like much on it's own, but combined with improvements that AMD are expected to implement before they release Bulldozer's successor, it could give them an edge in some markets.