[Maui, United States = DigitalToday reporter Dae-geon Seok] Qualcomm is expanding a Modular-based AI software stack across the Snapdragon family of products. Qualcomm views the growing number of AI hardware types as increasing the burden of having to build separate software for each chip. Chris Lattner (Chris Lattner), Qualcomm's senior vice president for advanced AI software and platforms, said in a Snapdragon Summit keynote on Sept. 23 local time that the Modular stack currently supports 6 architectures from 6 companies as a single stack.

Lattner cited software as one of the bottlenecks in the AI industry. The software stack refers to a bundle of languages, compilers and libraries needed to run AI models on a chip. "The industry is spending huge amounts of capital on diverse hardware, but the biggest problem was always software," he said. "Each hardware team had to build a stack from scratch."

Developers have had to deal with multiple different stacks despite similar functions. Bugs also vary by stack. AI models have also grown into a distributed-system problem with multiple layers intertwined, creating a structure in which falling short in even one layer leaves a player behind in competition.

To address this, Qualcomm acquired Modular in June. The purpose of the acquisition is to build an AI compute layer that is not tied to a specific chip across devices, the edge and data centers. Modular is a company Lattner co-founded and led as chief executive, and it has developed an integrated AI stack that runs across multiple hardware platforms over the past 4.5 years. Lattner joined Qualcomm with the acquisition.

The Modular stack consists of three layers: Mojo, MAX and Modular Cloud. At the bottom, Mojo is a high-performance programming language for AI hardware. Mojo keeps Python syntax while allowing direct hardware control, so the same task does not need to be written twice. Previously, most AI developers built models in Python, but Python does not run efficiently on hardware, requiring rewrites in lower-level languages such as C++ or CUDA to achieve performance. Modular released Mojo 1.0 last month and 공개했다 its source code under the Apache 2.0 license.

MAX, the layer above Mojo, is a framework for designing, building and serving AI models. Serving refers to operating a trained model so it responds to user requests. MAX helps run researcher-built models at high performance on target chips and works in both data centers and local PCs. It handles both the model-building stage and the execution stage within the same stack.

At the top, Modular Cloud is a layer for AI deployment companies that use large-scale compute resources. Modular handles server provisioning, scheduling, scaling, optimization and utilization management, while users only need to work with familiar APIs and tools. Modular recently expanded support from GPUs to data center accelerators and supports 6 architectures. That means code does not need to be rewritten for each accelerator. Qualcomm's data center accelerator is one of them, and Modular placed its stack on Qualcomm's NPU software and applied it to the accelerator.

Microsoft is working with Qualcomm's Modular team to support Mojo natively on Windows. Microsoft's goal is to let Windows developers do system-level programming with familiar Python syntax while achieving optimal performance.

Qualcomm will also expand the Modular stack to product lines outside data centers. "Qualcomm has a platform that spans from microwatts to kilowatts and beyond," Lattner said. "We are bringing the entire Modular software platform across Qualcomm's full portfolio, and that includes Snapdragon."