Senior NPU Compiler Engineer

Role: Senior NPU Compiler EngineerYears: 7-10 years | ChennaiBuild the compiler that brings Processing-in-Memory AI to life.InCommon is hiring on behalf of a California based company that is building ultra-low-power neural processing technology for edge AI — chips that run advanced AI workloads within tight power and memory budgets. The Processing-in-Memory (PIM) architecture computes directly inside memory, breaking the traditional bottleneck of shuttling data back and forth.We're looking for a Senior NPU Compiler Engineer to build the compiler stack — largely from the ground up — that takes models from TFLite, ONNX, and PyTorch and turns them into instructions our NPU can execute.What you'll do:🔹 Design and build the full compilation pipeline — graph import, operator lowering, scheduling, memory planning, and instruction generation🔹 Map neural network graphs onto a novel dataflow / Processing-in-Memory architecture🔹 Optimize for execution efficiency, memory bandwidth, and power🔹 Build validation tools to compare compiler output against simulator and RTL results🔹 Work directly with architecture, RTL, firmware, and ML teams to shape the hardware/software interfaceWhat you bring:✅ Strong experience building compilers, graph compilers, or backend toolchains✅ Solid C++ and Python✅ Deep understanding of compiler fundamentals — IRs, graph transforms, lowering, scheduling, codegen✅ Experience with ML model formats (ONNX, TFLite, PyTorch) and neural network operators/tensors✅ Background building software for NPUs, DSPs, GPUs, or other specialized acceleratorsNice to have:⭐ Experience with MLIR, LLVM, TVM, XLA, Glow, or IREE⭐ Familiarity with dataflow or Processing-in-Memory architectures⭐ Experience with binary instruction streams, DMA engines, or on-chip SRAMWhy join us:This is a rare chance to build a compiler stack from scratch for genuinely novel silicon — with real architectural influence, not just downstream implementation work. You'll shape decisions that affect the hardware itself, not just work around it.