SciSci Research is an AI chip design company with a novel architecture for radically faster and stronger arithmetic, the kind needed to drastically scale and improve AI.

 

SciSci's architecture replaces floating-point and fixed-point arithmetic with efficient, errorless compute. Based on some modern number theory (p-adic analysis), SciSci has a new arithmetic logic design and number representation format that enables exact arithmetic, without approximation or rounding error. By eliminating the inefficiencies and limitations of the decimal-based approximation inherent to floating-point and fixed-point, SciSci's architecture is expected to ​deliver speeds that are 15-50× faster than floating-point on average, and a 50× energy efficiency gain.

​

In addition to providing better speed and efficiency, exact arithmetic offers a more consistent and interpretable way to implement AI models, which will be important for alignment and safety.

​​​

SciSci's arithmetic logic design and number representation format can be licensed and used for matrix multiplication units, tensor cores, and other units used in AI chips.

​​