AI and ML
Developers await open source compiler release to dispel uncertainty following Qualcomm acquisition
Modular's Mojo reached its 1.0 milestone on Tuesday, promising developers a stable, production-ready, general-purpose programming language.But its primary purpose is more specific: to unify the AI software stack by providing a way to target the diverse set of GPUs, CPUs, ASICs, and other heterogeneous systems that are used to run AI workloads.Mojo, in conjunction with the MAX inference framework, aspires to make it easier to develop AI software, systems, and tools without being tied to vendor-specific platforms like Nvidia's CUDA (Compute Unified Device Architecture) or AMD's ROCm.
CUDA is Nvidia's proprietary platform for programming GPUs and works with the company's own hardware. Many AI-oriented tech companies that compete with Nvidia would prefer an open source alternative that can be optimized for their particular wares.
Chris Lattner, creator of LLVM, the Clang compiler, the Swift programming language, and the MLIR compiler infrastructure, has been working on Mojo publicly since 2023. He'd been doing so through a startup he co-founded called Modular that was acquired by Qualcomm in June. Last year, in a presentation, Lattner acknowledged the debt the AI industry owes to CUDA, but said the world has moved on and CUDA hasn't.Mojo brings the readability and accessibility of Python – the de facto language of AI – together with Rust-like memory safety and system programming capabilities that might otherwise require the use of system languages like C, C++, and CUDA.The language can utilize MLIR (Multi-Level Intermediate Representation Overview) to access AI-specific hardware features, thereby improving performance and optimizing compute costs. The 1.0 release brings with it several new features, including support for Python-style "lambda" syntax for inline closures, an improved Mojo Language Server Protocol (LSP) server for integrating with code editors like VS Code, and Mojo AI skills.
