Nova MLIR compiler hits 10.6% higher throughput than PyTorch

Nova, an automated end-to-end JIT compiler built on MLIR, tackles a core bottleneck in deep learning: mapping high-level tensor operations efficiently to physical hardware. Traditional eager-execution frameworks like PyTorch lack whole-graph visibility, forcing developers to hand-tune kernels or rely on coarse-grained optimizations. Nova captures eager executions and unifies forward and backward passes into a single value-semantic dialect, enabling aggressive whole-graph optimization across operation boundaries, complex memory hierarchies, and even register-level tuning.

The compiler uses an Analytic Configurator to deterministically derive optimal execution schedules from arithmetic intensity, reducing search time to zero. A structural hashing runtime then synthesizes fine-grained kernels directly from computation structure. On an RTX 3060, Nova matches or slightly beats cuBLAS and XLA on TF32 matmuls across most shapes, keeping relative error under 5e-4. At the model level, it achieves up to 10.6% greater throughput than PyTorch and 4.4% greater than XLA on a 42M-parameter model. Crucially, Nova cuts memory footprint by up to 29% versus PyTorch, allowing it to train a 144M-parameter model at 17,900 tokens/s on a 12GB GPU—a workload where PyTorch runs out of memory. This points toward a future where compiler-driven optimization is as important as model architecture in pushing hardware limits.