Senior AI Performance Modeling Architect

Advanced Micro Devices, Inc•Austin, TX

12h•Onsite

About The Position

As a Senior Engineer, you will lead engineers in developing the next generation of AI infrastructure by defining GPU architecture specifications that enable large-scale model training. You will drive performance improvements in training and inference pipelines through system design and optimization, champion advanced techniques across the engineering organization, and establish best practices for distributed ML systems that scale from one GPU to thousands. The role focuses on performance modeling and optimization for very large language-model training and inference, including dense and mixture-of-experts models across text, vision, and speech. You will develop parallelization strategies, architect memory-efficient training systems, incorporate advanced models and architectures, and collaborate with internal and external stakeholders and ML researchers.

Requirements

Extensive senior-level experience optimizing large-scale ML systems and GPU architectures.
Deep expertise in CUDA programming, GPU memory hierarchies, and hardware-specific optimizations.
Proven track record architecting distributed training systems for large-scale systems.
Expert knowledge of transformer architectures, attention mechanisms, and model-parallelism techniques.
Bachelor’s, MS, or PhD in Computer Science or Engineering, or equivalent industry experience.

Nice To Haves

Experience with PyTorch, CUDA, TensorRT, or OpenAI Triton.
Experience with distributed systems such as Ray or Megatron-LM.
Experience with performance analysis tools such as NSight Compute, nvprof, or PyTorch Profiler.
Experience with KV cache optimization, Flash Attention, or Mixture of Experts.
Experience with high-speed networking technologies such as InfiniBand, RDMA, or NVLink.

Responsibilities

Lead performance modeling and optimization for multi-trillion-parameter LLM training and inference, including dense and mixture-of-experts models with text, vision, and speech modalities.
Model and optimize parallelization strategies across tensor, pipeline, context, expert, and data-parallel dimensions.
Architect memory-efficient training systems using techniques such as structured pruning, quantization (MX formats), continuous batching, chunked prefill, and speculative decoding.
Incorporate and extend state-of-the-art models such as GPT-4, reasoning models such as Deepseek-R1, and multimodal architectures.
Collaborate with internal and external stakeholders and ML researchers to share results and iterate rapidly.