Highlights:

  • CUDA is the software interface through which developers engage with Nvidia’s chips.
  • Intel enhanced the framework with SYCLomatic, converting CUDA software into SYCL for other AI chips.

The UXL Foundation, comprised of leading players in the chip industry, anticipates that its alternative to Nvidia Corp.’s CUDA technology will reach a state of maturity by the end of the year.

Reuters recently released the group’s development timeline. The technology being developed by members of the UXL Foundation could streamline the migration of workloads from Nvidia chips to competing processors, potentially fostering increased competition for the graphics card manufacturer in the future.

The software interface used by developers to engage with Nvidia’s chips is referred to as CUDA. It furnishes the essential programming components required to execute a neural network on the company’s graphics cards. Moreover, CUDA incorporates tools that simplify associated duties, such as distributing an AI model across multiple GPUs.

Applications powered by CUDA can exclusively operate on Nvidia silicon. Adapting such software for AI accelerators produced by competing chipmakers is feasible but demands a substantial investment of time and effort. The UXL Foundation was established in September with the aim of facilitating this process.

Qualcomm Inc., Samsung Electronics Co. Ltd., Intel Corp., and a number of other significant chipmakers support the consortium. According to a report published by Reuters recently, the UXL Foundation is seeking to include more major players in the chip industry as members. It’s thought the group also wants to attract big cloud providers.

Up to this point, the development endeavors of the UXL Group have been centered around a software toolkit named OneAPI. Originating from Intel, this toolkit facilitates the creation of AI applications that can be transferred between various chips with relative simplicity. OneAPI builds upon a preceding framework known as SYCL, which similarly prioritizes the facilitation of application portability.

Intel has added a number of new features to the framework, the most notable of which is a function known as SYCLomatic. Its purpose is to translate software meant for Nvidia’s CUDA into SYCL code that can be executed on AI chips made by other companies. “We’re actually showing developers how you migrate out from an Nvidia platform,” Vinesh Sukumar, Qualcomm’s AI and machine learning lead, informed Reuters.

By minimizing the effort and, consequently, the expenses associated with transitioning Nvidia-powered applications to competitor chips, there is a possibility of increasing the likelihood of enterprises adopting those chips. Over time, this could foster heightened competition for Nvidia’s dominant market position in graphics cards.

According to Reuters, oneAPI is “already usable.” According to reports, the UXL Foundation wants to advance the technology in order to develop a “standard programming model of computing designed for AI.”

Thus far, the endeavor has attracted technical input from both the consortium’s member firms and external entities. It is anticipated that the technical steering committee of the UXL Foundation will finalize the specifications for its AI programming model in the first half of 2024. The technology is projected to reach maturity by the end of the year.