About the Role
We are seeking a highly experienced Compiler / Software Architect to join our NPU Hardware Architecture team and play a key role in the hardware-software co-design of next-generation AI inference accelerators.
This is a unique role at the intersection of NPU hardware architecture, AI compilers, performance modeling, and software development. The position will work directly with hardware architects to define, evaluate, and prototype the software stack required to program current and future NPU architectures.
Our NPU architecture is closely coupled with the compiler stack. Efficient utilization of the accelerator requires deep understanding of the hardware compute architecture, memory hierarchy, dataflow, scheduling, quantization, instruction set and execution model. The successful candidate will use this understanding to develop compiler and software proof-of-concepts, evaluate architectural proposals against real AI workloads, and influence the design of future NPU hardware.
You will work closely with hardware architects, micro-architects, compiler engineers and AI software teams to answer a fundamental question:
How should the NPU hardware and software be co-designed to deliver the best performance, power efficiency, programmability and scalability for real-world AI workloads?
What You Will Be Responsible For
Compiler & Software Architecture for NPU
- Develop compiler and software proof-of-concepts for current and next-generation NPU architectures.
- Define software abstractions and compiler flows that efficiently expose NPU hardware capabilities to AI workloads.
- Develop and evaluate compiler concepts including graph lowering, intermediate representations, operator mapping, scheduling, tiling, fusion, memory planning and code generation.
- Translate NPU architectural concepts into executable software models and demonstrate their feasibility using representative AI workloads.
- Develop lightweight compiler/runtime infrastructure to validate new hardware features before production software implementation.
- Analyze existing compiler limitations and identify architectural changes required to improve programmability and accelerator utilization.
Hardware–Software Co-Design
- Work as an integral member of the hardware architecture team to co-design NPU hardware and software.
- Analyze how proposed hardware features can be effectively exposed through the compiler and software stack.
- Provide software-driven feedback on compute architecture, memory hierarchy, data movement, dataflow, scheduling, instruction set and accelerator programmability.
- Identify hardware features that provide meaningful benefits to real AI workloads and challenge features that add hardware complexity without sufficient software value.
- Define compiler requirements and software abstractions for new NPU capabilities.
- Participate in architecture and micro-architecture reviews and influence hardware decisions from a software and workload perspective.
- Evaluate architectural trade-offs considering performance, power, area, compiler complexity and software scalability.
AI Workload & Performance Analysis
- Analyze representative AI models and workloads to identify compute, memory, bandwidth, scheduling and data-movement bottlenecks.
- Build software-based performance models and workload prototypes to evaluate architectural concepts.
- Develop experiments to quantify the impact of proposed hardware features on model performance and accelerator utilization.
- Investigate issues such as quantization, sparsity, operator fusion, tensor layouts, tiling, data reuse, memory bandwidth and scheduling efficiency.
- Correlate software/model-level performance with architectural and micro-architectural behavior.
- Use workload analysis to guide both current-generation optimizations and next-generation NPU architecture.
Architecture Prototyping
- Rapidly prototype software solutions for architectural concepts that may be months or years away from production silicon.
- Develop functional models, compiler prototypes, simulators, emulators, reference implementations or runtime abstractions as needed to validate architectural ideas.
- Demonstrate end-to-end execution of representative AI workloads on proposed NPU architectures.
- Build proof-of-concepts that allow hardware architects to make informed architectural decisions before RTL implementation.
- Help establish software models and interfaces that can later evolve into production compiler components.
Cross-Functional Leadership
- Work closely with NPU hardware architects, micro-architects, RTL designers and the production compiler/software organization.
- Bridge the communication gap between hardware and software teams and translate requirements in both directions.
- Participate in architecture definition from early concept through implementation and silicon bring-up.
- Mentor engineers and contribute to technical direction for compiler-driven hardware/software co-design.
- Influence the roadmap of future NPU architectures through workload and software-driven insights.
Necessary Qualifications
Bachelor's or Master's degree in Computer Science, Computer Engineering, Electrical Engineering or a related field.
8+ years of experience in compiler development, systems software, computer architecture, AI accelerators, or a closely related area.
Strong understanding of compiler architecture and compiler optimization techniques.
Strong C++ programming skills and proficiency in Python.
Strong understanding of computer architecture and familiarity with accelerator architectures, memory hierarchies and data movement.
Experience analyzing AI/ML workloads and understanding performance bottlenecks.
Ability to work across hardware and software boundaries and communicate effectively with hardware architects.
Strong analytical and problem-solving skills, with the ability to rapidly prototype and evaluate architectural ideas.
Good to have
- Experience developing compilers or software stacks for NPUs, GPUs, DSPs, TPUs or other AI accelerators.
- Experience with MLIR, LLVM, TOSA, StableHLO, Torch-MLIR or similar compiler infrastructure.
- Experience with accelerator-specific scheduling, tiling, memory management or code generation.
- Understanding of NPU architecture concepts such as dataflow, tensor engines, systolic/array-based compute, local SRAMs, DMA/data movement and accelerator instruction sets.
More information about NXP in India...
#LI-2734