Intel has published a nine-page long whitepaper that discusses the game physics performance of the Larrabee architecture, you can read it over here.
Game physics is at the heart of any modern game engine which employs the laws of physics to simulate life-like movement and interaction between objects, such as rigid and deformable bodies, cloth, and water. Game physics applications are very compute and memory intensive. The ever growing quest for a high degree of realism requires more complex physics algorithms and also larger datasets. To meet the demands of game physics applications requires a computer architecture which can deliver high floating point performance and memory bandwidth. However, general-purpose many-core architectures are quickly evolving to overcome these constraints. Larrabee is one such highly-threaded many-core architecture. It consists of an array of multiple IA Intel processor cores, each augmented with a 16-wide vector processor unit. In this paper, we analyze several key game physics applications. We show how Larrabee’s extensive thread and data parallelism are well suited for a broad set of physics algorithms. We show how these algorithms parallelize and map to Larrabee architecture and achieve good parallel speedup with the number of cores.