AI has pushed an explosion of recent quantity codecs—the methods by which numbers are represented digitally. Engineers are each potential strategy to save computation time and power, together with shortening the variety of bits used to characterize knowledge. 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 lately joined Barcelona-based Openchip as an AI engineer, about his efforts to develop a bespoke quantity format for scientific computing.
LASLO HUNHOLD
Laslo Hunhold is a senior AI accelerator engineer at Barcelona-based startup Openchip. He lately accomplished a Ph.D. in pc science from the College of Cologne, in Germany.
What makes quantity codecs attention-grabbing to you?
Laslo Hunhold: I don’t know one other instance of a area that so few are inquisitive about however has such a excessive affect. For those who make a quantity format that’s 10 p.c extra [energy] environment friendly, it could translate to all functions being 10 p.c extra environment friendly, and it can save you plenty of power.
Why are there so many new quantity codecs?
Hunhold: For many years, pc customers had it very easy. They might simply purchase new techniques each few years, and they’d have efficiency advantages without spending a dime. However this hasn’t been the case for the final 10 years. In computer systems, you have got a sure variety of bits used to characterize a single quantity, and for years the default was 64 bits. And for AI, firms observed that they don’t want 64 bits for every quantity. So they’d a robust incentive to go all the way down to 16, 8, and even 2 bits [to save energy]. The issue is, the dominating commonplace for representing numbers in 64 bits isn’t 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 giant numbers, or very small numbers, and really excessive accuracy in each instances. The 64-bit commonplace 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 observe a selected distribution, and also you don’t want as a lot accuracy.
What makes a quantity format “good”?
Hunhold: You’ve got infinite numbers however solely finite bit representations. So you must determine the way you assign numbers. A very powerful half is to characterize numbers that you just’re really going to make use of. As a result of when you characterize a quantity that you just don’t use, you’ve wasted a illustration. The best factor to have a look at is the dynamic vary. The following is distribution: How do you assign your bits to sure values? Do you have got 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 steadily could be represented with extra density. However posits don’t work for scientific computing, and this can be a enormous difficulty. 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. Folks 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 employ in scientific computations, when you have a look at all of the fields, and designed takums such that if you take away bits, you don’t scale back that dynamic vary
This text seems within the March 2026 print difficulty as “Laslo Hunhold.”
From Your Web site Articles
Associated Articles Across the Internet
