My current research focuses on emerging floating-point numbers. In particular, my work tries to develop a hardware platform using RISC-V CPUs to explore posit arithmetic in domain-specific applications.

Exploring new arithmetics is a way of optimizing the execution of specialized compute-intensive applications. This is the case of machine learning, for example with the bfloat16 format present in TPUs, or scientific computing.

I am researching hardware implementations of posit arithmetic. This is a promising alternative to IEEE 754 floats and doubles (the ones we are all used to) that was proposed in 2017. Posits have a good trade-off between dynamic range and accuracy and encounter few exceptions when operating (in fact, only two: \(0\) and \(\pm\infty\)). They also have tapered precision, which means that numbers near \(\pm 1\) are more accurate, while very big and very small numbers have less accuracy.

You can find all the details in my publications.