Former Apple Luminary Sets Out to Create the Ultimate GPU Software (www.wired.com)

Chris Lattner — creator of LLVM and Swift and a former Apple, Google and Tesla engineer — has launched Modular, a startup building a unifying software layer and Python-based language that lets developers run AI workloads across diverse GPUs and CPUs without rewriting code for each vendor. Modular, founded in 2022 with Tim Davis, just raised $250M in a third funding round (valuation $1.6B) and says its platform now supports Nvidia, AMD and Apple Silicon GPUs; partners include Nvidia, AMD and Amazon. The product aims to close the gap between proprietary stacks like Nvidia’s CUDA and alternatives such as AMD’s ROCm by shipping optimized kernels (e.g., attention kernels) faster and presenting a single programming model for heterogeneous hardware. This matters because GPU software fragmentation is a major bottleneck for scaling generative AI: CUDA’s dominance locks developers in, ROCm is fragmented by vendor, and writing GPU kernels is “dark art” requiring rare expertise. Modular’s approach — a compiler/runtime and language layer that automates optimization across chips and complements existing toolchains — could make cloud deployments more portable, accelerate adoption of non‑Nvidia silicon, and lower engineering costs. Challenges remain: potential vendor resistance, customer reluctance to pay for an extra layer, and competition from startups (like Mako) using AI to auto-generate optimized kernels rather than a new universal compiler.