Dr. Dobb's | Intel Sheds Light on Larrabee Processor

Intel Sheds Light on Larrabee Processor

The chipmaker reveals details of the upcoming processor, which will feature multiple x86 cores, and will be fully programmable.

August 04, 2008
URL:http://www.drdobbs.com/architecture-and-design/intel-sheds-light-on-larrabee-processor/209902667

Intel has fired a shot across NVIDIA's bow by revealing some details of its upcoming Larrabee processor at a briefing Friday in San Francisco.

Larrabee will initially take life as a graphics processor supporting DirectX and OpenGL, but is also designed with parallelization of non-graphics tasks in mind, and can form the basis of highly parallel add-in processor hardware akin to NVIDIA's Tesla products.

With Larrabee, Intel is playing to its strengths by designing a CPU-based, rather than GPU-based solution. Larrabee will consist of multiple cores, each of which is a complete x86 core, with everything that that entails: context switching and pre-emptive multitasking, virtual memory, and cache coherency.

Larrabee will be fully programmable, and will contain only one fixed-function processing stage: a texture sampler. In traditional GPUs, rendering pipelines contain multiple fixed-function stages that must be passed through. In contrast, every step in the Larrabee pipeline is essentially software-controlled, allowing for highly customized rendering.

[Click image to view at full size]

Larrabee block diagram.

Larrabee's L2 cache is partitioned among the cores, which communicate via a ring bus designed provide fast synchronization between cores and caches.

While Intel spokespersons were cagey about product details, Stephen Junkins of Intel's Software Solutions Group said "The first products will look like a GPU, but future products? Who knows?" First products based on Larrabee are expected in 2009 or 2010.

Alongside the DirectX and OpenGL API, Larrabee will support its own native C/C++ API for customized processing not supported by the other APIs. Intel representatives acknowledged the possibility for codebase forking that this introduces, but emphasized that they wanted to give developers the flexibility to code custom algorithms, and hoped that forking could be kept to a minimum through the use of middleware. The native C/C++ API will also allow for the coding of a range of scientific and high-performance applications that will take advantage of the many-core Larrabee architecture.

Intel will present a paper on Larrabee at SIGGRAPH in Los Angeles on August 12th.