What about TVM, XLA, and AI compilers? (www.modular.com)

As AI models have grown in size and complexity, manually writing optimized GPU kernels has become impractical, pushing the AI community toward automated AI compilers that transform high-level operations into efficient hardware-specific code. Projects like TVM and Google’s XLA exemplify this shift, enabling kernel fusion—combining operations like matrix multiplication and ReLU activation into a single GPU kernel to dramatically reduce memory overhead and improve performance by up to 2x. These compilers address the explosive complexity caused by thousands of operators, new numeric formats like float8, and evolving hardware targets. TVM, an early open-source AI compiler developed from academic roots, made strides in cross-platform optimization and inspired broad industry adoption. However, it struggled to keep pace with new specialized hardware such as NVIDIA's Tensor Cores and GenAI workloads, suffered from fragmentation due to vendor forks, and faced slow compile times. Meanwhile, Google’s XLA—originally designed to optimize TPU performance—benefited from vast engineering resources and became a foundation for large-scale model training and distributed computation. Yet, XLA’s dual identity as a proprietary TPU compiler and a less actively maintained open-source GPU project (OpenXLA) has limited its impact outside Google’s TPU ecosystem. This history highlights a fundamental tension in AI compiler development: balancing peak hardware-specific performance with generality, ease of evolution, and ecosystem governance. As GenAI fuels soaring compute demand across diverse platforms, the community urgently needs compilers that combine agility, broad hardware support, and strong industry collaboration to keep pace with rapid AI innovation.