Specifications

April 2012 v1 AMD Opteron™ 6200 Linux Tuning Guide
5
The AMD Opteron™ 6200 Series processors can boost core frequencies by allocating more power to individual
cores, subject to a total power limit and other restrictions. Cores that are idle can “go to sleep” and turn off their
power draw, allowing more power to be dedicated to active cores. As a result, a single core job could see a
processor running at 3.2Ghz, which all cores active see cores running at 2.3Ghz (for top 16-core standard power
AMD Opteron™ 6276 Series processors).
This following block diagram of the G34 socket Interlagos processor shows a 16-core variant of the part. There
are a few new key features that have a significant effect on performance. As an introduction to this topic, note
that the device is organized as a Multi-Chip Module, which means that there are two CPU dies in the package,
interconnected using a HyperTransport link. Each die has its own memory controller interface to allow external
connection of up to two memory channels per die.
This memory interface is shared among all the CPU core pairs on each die. Each die is organized as 2-, 3-, or
4-core pairs, and each core pair has two CPU cores, each of which has a set dedicated integer unit and general
purpose register files. In addition, there are resources that are shared between the cores in a pair that operate
at double rate, including instruction cache, instruction decode logic, Floating Point Unit, and L2 cache. Each core
pair appears to the operating system as two completely independent CPU cores. Thus, to the OS, the device
shown in the following block diagram appears as 16 CPUs.
Sharing of CPU resources is an engineering design decision intended to provide superior performance for a
broad range of applications while maximizing power efficiency. Since most of the shared resources can provide
more performance than a single CPU core can take advantage of, it makes good engineering sense to share
among multiple cores.
The fact that some resources are shared has implications for performance since, in some cases, the shared
resources may not satisfy the full demand from an application running on both CPU cores. Because of this, it is
important to understand more about these shared resources.