Centreon develops cutting-edge integer and floating-point computational units for use in AI, high-performance computing, and scientific computing applications. Compared to commonly-available commercial designs for floating-point adders and multipliers, we deliver comparable-to-superior performance and features (such as all-in-one units and in-path subnormal handling) using up to 39% fewer transistors, allowing computation-dense chips to pack more arithmetic units in their layout and achieve greater throughput. We know that in a computation-dense chip like a specialized AI chip, vast amounts of chip area are consumed by duplicated adders and multipliers; our units can handle multiple levels of precision and handle both integer and floating-point results with the same hardware, reducing the need for redundant units by up to 66%. We also provide fully-pipelined implementations of nonlinear functions such as the trigonometric functions, eliminating the traditional need for expensive memory accesses and lookup tables in computing such functions. Our units have been used in silicon-realized, commercial chips, such as in FuriosaAI’s Tensor Contraction Processor’s vector engine.

Parkson Wong received an M.S. in Computer Science from the University of Wisconsin-Madison, and went on to work as a Principal Scientist at the NASA Ames Research center, where he conducted research in high-end and heterogeneous computing. In the same period, he was the Principal Engineer at Silicon Graphics Inc., where he implemented and deployed high-definition video servers for major broadcast facilitating. His latest pursuit is as co-founder of Centreon Technology, where he brings four decades of computer architecture and systems design experience to achieve unparalleled performance.

Benjamin Ou received a B.A. in Computer Science and Physics from the University of California, Berkeley, during which he interned at Centreon Technology. He went on to accept a full time position with Centreon, where he initially worked on benchmarking cutting-edge hardware using standard benchmarking libraries as well as the GROMACS molecular simulation platform. Currently, he along with Parkson leads Centreon’s effort to design its own vector-capable out-of-order CPU that aims to bring the most out of Centreon’s unparalleled computational units.