Join us as we welcome Theo Drane, a Fellow silicon design engineer at AMD. He'll be presenting "Novel Numerical Hardware Design Methodology: From Machine Readable Specification to Optimized RTL."
ABSTRACT: Machine learning has given rise to a plethora of new number formats, combined with the hundreds already present within GPUs. This has triggered new challenges in specifying, implementing and verifying associated numerical hardware. Architects need to specify, simulate and explore a whole range of number formats and precisions, which is not easily done within a standard modelling language such as C/C++.
Theo Drane presents a language that can represent any number format and precision in its purest form, which can be simulated against an infinitely precise reference and used to automatically
explore a multi precision design space. Once the bit accurate intent is determined, he provides a
tool chain that automatically generates a bit accurate C-model.
Equivalently, a bit identical RTL is also generated. Such behavioral RTL typically results in very poor power, performance and area. Going beyond the capabilities of current industrial high level synthesis tools, he will also present a tool specifically targeted at optimizing such Datapath designs capable of automatically generating competitive optimized RTL. The specification language and tool chains have been built on a foundation of formal verification, which provides visibility and trust of the methodology.
THEO DRANE is a fellow silicon design engineer at AMD. He worked for the Datapath consultancy, Arithmatica, after completing a mathematics degree from the University of Cambridge. Moving to Imagination Technologies, he founded their Datapath team while studying for a doctoral degree at Imperial College London. After a stint within Cadence Design System’s logic synthesis division, Genus, he joined Intel’s 3D compute and graphics division where he founded the numerical hardware and system level design group. During these positions, he has created applied research groups which act as internal consultancies focusing on all aspects of architecting, implementing, optimizing and formally verifying math intensive hardware and system level design in general. He promotes the ‘acadustrial’ doctorate model: simultaneous full-time industrial employment and doctoral degree enrollment, driving knowledge creation and transfer directly into production.
Read More