AMD exposes Bulldozer and Bobcat cores

Attempts to better Hyper-Threading

With Fusion's grand entrance rapidly approaching, AMD has finally decided to stop being all secretive and start dishing out the details.

The company is revealing all the juicy gossip about its forthcoming Bobcat and Bulldozer cores at the Hot Chips symposium today, but it also kindly shared the information with us beforehand.

As you may already know, Fusion sees an entirely new approach to building CPUs, where a processor is constructed from modules. These modules can be anything from a CPU core, to a video decoding unit to an array of GPU stream processors. Bulldozer and Bobcat are specifically the CPU core modules, but these also show a very different design from AMD's previous CPU cores.

Let's start with Bulldozer, which as its name suggests, is the heavyweight core for servers and desktops. This 32nm silicon-on-insulator (SOI) core sees AMD finally taking on Intel's Hyper-Threading technology, but with a very different attitude.

Rather than executing two threads on a single core, where the threads will be jostling for resources, Bulldozer instead splits the load over two integer units. According to AMD, this is based on the principle that '80 per cent' of compute tasks use integer (whole number) calculations.

Each integer unit has its own allocation of Level 1 cache, but shares a block of Level 2 cache and a large floating point (for calculations with decimal points) unit with the other integer blocks. In addition to this, all the Bulldozer modules in a Fusion chip will have access to a shared pool of Level 3 cache at the chip level, as well as the memory controller.

Basically, you get nearly two whole processing cores inside a single Bulldozer module, which should enable it to crunch through two threads much more convincingly than an Intel core with Hyper-Threading. Whether it works in practice remains to be seen, but it's an efficient piece of design in theory.

According to AMD, Bulldozer dynamically switches between its shared and dedicated components in order to maximise the performance per watt, making it much more efficient in terms of both speed and power consumption. Several Bulldozer modules can then be chained together in a single CPU package, meaning a quad-core CPU can effectively execute eight simultaneous threads.

Bobcat on the next page...