Senior Researcher - Edge AI Optimization/Hardware-Aware ML

Huawei Canada has an immediate permanent opening for a Researcher.

About the team:

The Software-Hardware System Optimization Lab focuses on research and innovation in power efficiency and performance optimization for consumer devices. By leveraging the talents and capabilities of local academia and our team, we aim to build system-optimization capabilities for software and hardware across edge AI, multimedia, graphics, mobile gaming, and system software domains, thereby enhancing the user experience and performance competitiveness of Huawei's consumer device products.

About the job:

• Conduct research in hardware-aware neural network optimization (e.g., quantization-aware training, mixed precision, pruning, distillation, neural architecture search).
• Develop novel approaches for latency/energy-aware training objectives and Pareto optimization (accuracy vs. compute vs. memory).
• Prototype and evaluate techniques for efficient inference under device constraints (thermal limits, memory bandwidth, intermittent connectivity).
• Publish and present findings internally and externally (papers, workshops, patents, technical blogs).
• Optimize inference pipelines across pre/post-processing, scheduling, operator fusion, memory planning, and runtime execution.
• Collaborate on or contribute to compilers / runtimes (e.g., TVM, MLIR, XLA, TensorRT, ONNX Runtime, TFLite, ExecuTorch) to improve operator coverage and performance.
• Profile and optimize models with real device traces, addressing bottlenecks such as cache misses, memory bandwidth, kernel launch overhead, and CPU–NPU handoff.
• Build and maintain hardware-aware benchmarking methodology and regression suites for edge targets (ARM CPU, mobile GPU, DSP, NPU).
• Create deployment recipes for heterogeneous compute (CPU+GPU+NPU) including partitioning strategies and fallback paths.
• Drive optimization for on-device personalization and incremental updates when needed (e.g., small adapters, efficient fine-tuning).
• Partner with product engineering, platform teams, and hardware teams to translate device constraints into research targets and to transition research prototypes into production.
• Mentor junior researchers/engineers, review experimental designs, and raise the quality bar for measurement rigor and reproducibility.
• Define technical roadmap areas (e.g., next-gen quantization, kernel optimization, model families for edge, compiler improvements).

About the ideal candidate:

• PhD (or equivalent research experience) in Machine Learning, Computer Science, Electrical/Computer Engineering, or related field.
• Strong programming skills in Python and C/C++ (or equivalent systems language).Experience building AI agent / harness / skill toolchains, including model evaluation, orchestration, and LLM-powered tooling. Hands-on experience with deep learning frameworks (e.g., PyTorch, TensorFlow, JAX) and deployment toolchains (e.g., ONNX, TFLite, TensorRT, TVM, MLIR-based stacks). Solid knowledge of performance profiling: latency measurement, memory profiling, kernel-level bottleneck analysis, and experimental rigor.
• Proven publication record at top venues (e.g., NeurIPS/ICML/ICLR, MLSys, ASPLOS, ISCA, MICRO) and/or patents in ML efficiency.
• 2+ years of relevant experience (research lab or industry) with demonstrated impact in at least one of:
• Model compression (quantization/pruning/distillation)
• Efficient architectures (MobileNet-like, MoE, efficient transformers, etc.)
• ML systems/compilers/runtime optimization
• hardware-aware optimization for edge deployment
• Experience optimizing for specific edge hardware:
• ARM NEON, mobile GPUs, DSPs, NPUs, microcontrollers
• Experience with distributed benchmarking, CI for performance regression, and reproducible experiment pipelines.
• Understanding of power/thermal constraints and methodologies for measuring energy on device.
• Experience with efficient LLM/VLM inference on edge (KV-cache optimization, quantized attention, speculative decoding, etc.).
• Technical Skills：
• Quantization: PTQ/QAT, per-channel/per-tensor, calibration, smooth quant, GPTQ-like methods, mixed precision
• Sparsity: structured pruning, N: M sparsity, hardware-friendly sparsity
• Compiler techniques: graph rewriting, operator lowering, scheduling, kernel autotuning
• Runtime techniques: memory arenas, tensor lifetime analysis, static vs dynamic shapes, batching strategies
• Hardware fundamentals: cache hierarchy, SIMD, memory bandwidth, accelerator programming models