Research Engineer: AI Workload & Systems

Huawei Technologies Canada Co., Ltd.; Markham; Sur place; CAD 178 070 - 315 479; Plein temps

Résumé du poste

Huawei Canada is seeking a Researcher for a 12-month contract to join the Frontier AI Technology Research initiative. The role focuses on hardware-aware AI workload analysis, kernel optimization, and evaluating next-generation accelerators for large-scale model training and inference.

The successful candidate will track evolving AI architectures (LLMs, MoE, sparse/linear attention), develop performance models, and provide architecture guidance to improve data efficiency and system performance in

Qualifications

• Experience with AI accelerator architectures and memory hierarchies.
• Hands-on kernel development using CUDA, Triton, or TileLang.
• Experience with LLM inference systems and serving frameworks like vLLM or SGLang.
• Ability to profile with Roofline-based methodologies.
• Publication or open-source contributions in AI infrastructure are a plus.
• Ph.D. in AI/CS is an asset.

Responsabilités

• Track evolution of state-of-the-art AI model architectures including LLMs, VLMs, MoE.
• Analyze computational characteristics of model architectures and workloads.
• Develop a framework to map AI workloads to hardware requirements.
• Identify patterns, design efficient inference deployment strategies, build performance models.
• Provide architectural recommendations for AI accelerators.

Connaissances

AI architectures

AI accelerators

Kernel development

CUDA

Triton

PyTorch

JAX

Roofline Profiling

MoE

Formation

Ph.D. in AI / CS

Outils

CUDA

Triton

TileLang

PyTorch

JAX

Description du poste

Huawei Canada is seeking a Researcher for a 12-month contract to join the Frontier AI Technology Research initiative. The role focuses on hardware-aware AI workload analysis, kernel optimization, and evaluating next-generation accelerators for large-scale model training and inference.

The successful candidate will track evolving AI architectures (LLMs, MoE, sparse/linear attention), develop performance models, and provide architecture guidance to improve data efficiency and system performance in