Smoothing over the switch
Reaching the next frontier in high-accuracy and energy-efficient computing will depend not just on posit arithmetic but the underlying platforms that can support it. According to Gustafson, researchers around the world have already built math libraries that will support conversion routines between one data type to another, like floats to posits and vice versa.
At its current stage, posit representation functions as if it were the more widely used floats, but provides much improved calculations, even for complex operations such as exponents and trigonometric functions like sines. As part of the posit standard, which Gustafson is developing and soon set to publish officially, such conversion would pave the way for wider uptake of the posit format.
“Having a standard you can point to and having portability between all the different systems is key to having at least one major company embrace posits to the point where it’s supported on what they produce as a product,” he said.
Among the recent advancements in this space is PACoGen, created by University of Hong Kong researchers for generating posit arithmetic cores that can perform basic operations.
As these cores form part of the circuitry of computers, PACoGen serves as a pipeline for sketching out the integration of posit processing units (PPUs) into hardware like microchips, showing how posit formats can fit within various technical specifications.
Another crucial piece of the puzzle is posits’ compliance with RISC-V instruction set architecture, an open-source set of design specifications that ensure compatibility across different processors.
“The breakthrough was realizing I could give up on divides as a perfect operation. If I can make add, subtract and multiply perfect, then I don’t have to have divides be perfect because they’re less common,” explained Gustafson.
Like picking only the battles that can be won, this optimization led to a clean number system that fit in a neat size and presented the same ordering as integers.
While posits were conceptualized as drop-in replacements for floats, Gustafson acknowledged that there will be a transition period where users can switch between the two formats, akin to having different processing modes.
From the frontiers of computing to mainstream tech, this transitional phase has already been set in motion. At ThoughtWorks Technologies India, for example, scientists replaced floats with PPUs on their Rocket Chip.
Evaluated on the RISC-V test suite, they found that posits minimized the amount of unusable representations appearing in the calculations. At the same time, the PPUs still supported RISC-V floating point extensions—additional codes allowing certain functions to be carried out as needed.
As new processors are already being designed, Gustafson envisions that posits would become the dominant format supported by the hardware, while floats become more of a rarely used legacy tool. Floats would still be available, but only through software extensions and libraries. For him, overcoming intellectual inertia and making the switch to posits could happen as soon as the next decade.
“I think posits will gradually be adopted and they will compete, and therefore computing will move forward in the same way,” concluded Gustafson. “That kind of improvement is gradual and I don’t think it has to be that painful. I hope posits will eventually take over the world.”
This article was first published in the print version of Supercomputing Asia, July 2021.
Click here to subscribe to Asian Scientist Magazine in print.
———
Copyright: Asian Scientist Magazine. Illustration: Oikeat Lam/Asian Scientist Magazine.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.
