High-Performance Physical Simulations on Next-Generation Architecture with Many Cores.

Yen-Kuang Chen, Jatin Chhugani, Christopher J. Hughes, Daehyun Kim, Sanjeev Kumar, Victor Lee, Albert Lin, Anthony D. Nguyen, Eftychios Sifakis, Mikhail Smelyanskiy.

In the Intel Technology Journal, August 2007.

Physical simulation applications model and simulate complex natural phenomena. The computational complexity of real-time physical simulations far exceeds the capabilities of modern unicore microprocessors, which are limited to only tens of billions floating- point operations per second (FLOPS). However, the advent of multi-core architectures promises to soon make processors with trillions of FLOPS available. Such processors are also known as tera-scale processors. Physical simulations can exploit this huge increase in computational capability to increase realism, enable interactivity, and enrich a user's visual experience.

In this work, we study physical simulation applications in two broad categories: production physics and game physics. After parallelization, the benchmark applications achieve parallel scalabilities of 30x.60x on a simulated chip-multiprocessor with 64 cores.

We examine the memory requirements of physical simulation applications and find that they require cache sizes in excess of 128MB and main memory bandwidths in excess of hundreds of GB/s for real-time performance. A radical re-design of the memory hierarchy may be necessary for the multi-core tera-scale era to provide good scaling for this type of application.