Exploring Innovative Number Formats for AI Efficiency

AI has pushed an explosion of recent quantity codecs—the methods through which numbers are represented digitally. Engineers are taking a look at each doable approach to save computation time and energy, together with shortening the variety of bits used to symbolize information. However what works for AI doesn’t essentially work for scientific computing, be it for computational physics, biology, fluid dynamics, or engineering simulations. IEEE Spectrum spoke with Laslo Hunhold, who not too long ago joined Barcelona-based Openchip as an AI engineer, about his efforts to develop a bespoke quantity format for scientific computing.

What makes quantity codecs fascinating to you?

Laslo Hunhold: I don’t know one other instance of a area that so few are fascinated with however has such a excessive influence. For those who make a quantity format that’s 10 % extra [energy] environment friendly, it could actually translate to all purposes being 10 % extra environment friendly, and it can save you quite a lot of vitality.

Why are there so many new quantity codecs?

Hunhold: For many years, laptop customers had it very easy. They may simply purchase new programs each few years, and they might have efficiency advantages without spending a dime. However this hasn’t been the case for the final 10 years. In computer systems, you’ve gotten a sure variety of bits used to symbolize a single quantity, and for years the default was 64 bits. And for AI, corporations observed that they don’t want 64 bits for every quantity. So that they had a powerful incentive to go all the way down to 16, 8, and even 2 bits [to save energy]. The issue is, the dominating customary for representing numbers in 64 bits is just not nicely designed for decrease bit counts. So within the AI area, they got here up with new codecs that are extra tailor-made towards AI.

Why does AI want totally different quantity codecs than scientific computing?

Hunhold: Scientific computing wants excessive dynamic vary: You want very massive numbers, or very small numbers, and really excessive accuracy in each circumstances. The 64-bit customary has an extreme dynamic vary, and it’s many extra bits than you want more often than not. It’s totally different with AI. The numbers normally comply with a selected distribution, and also you don’t want as a lot accuracy.

What makes a quantity format “good”?

Hunhold: You’ve gotten infinite numbers however solely finite bit representations. So you should determine the way you assign numbers. An important half is to symbolize numbers that you just’re really going to make use of. As a result of for those who symbolize a quantity that you just don’t use, you’ve wasted a illustration. The only factor to take a look at is the dynamic vary. The following is distribution: How do you assign your bits to sure values? Do you’ve gotten a uniform distribution, or one thing else? There are infinite potentialities.

What motivated you to introduce the takum quantity format?

Hunhold: Takums are based mostly on posits. In posits, the numbers that get used extra regularly will be represented with extra density. However posits don’t work for scientific computing, and it is a enormous situation. They’ve a excessive density for [numbers close to one], which is nice for AI, however the density falls off sharply when you have a look at bigger or smaller values. Individuals have been proposing dozens of quantity codecs in the previous couple of years, however takums are the one quantity format that’s really tailor-made for scientific computing. I discovered the dynamic vary of values you utilize in scientific computations, for those who have a look at all of the fields, and designed takums such that once you take away bits, you don’t cut back that dynamic vary

This text seems within the March 2026 print situation as “Laslo Hunhold.”