AI Accelerator Software Distinguished Engineer- Framework Integration

About The Role

As an AI Accelerator Software Distinguished Engineer – Framework Integration, you will lead end-to-end technical strategy and delivery for high-performance deep learning inference across Ampere accelerator platforms. You will set direction for how major ML frameworks are enabled and optimized for our hardware, ensuring high throughput, low latency, and efficient compute/memory utilization for current and next-generation AI workloads spanning data centers to edge.

This role is distinguished by deep technical ownership, architecture leadership, and cross-team influence—driving outcomes from performance modeling and integration strategy through production-ready runtime and kernel behavior.

What You’ll Achieve

• Framework integration leadership (PyTorch / ONNX / llama.cpp)
• Own and advance integration of major deep learning frameworks—PyTorch, ONNX, llama.cpp, and related tooling—into the Ampere deep learning accelerator backend, enabling robust execution of real-world model graphs and operators.
• Full-stack acceleration across the SW/HW execution path
• Drive acceleration across the end-to-end stack, including (as applicable): inference serving and orchestration enablement, framework-to-runtime integration layers, compiler/graph lowering and optimization, runtime library and execution management, user-mode execution paths and performance-critical interfaces, compute kernel development and micro-optimizations, profiling, benchmarking, and continuous performance tuning
• Model enablement: performance + accuracy
• Lead efforts to improve performance and correctness for models using popular frameworks and serving stacks such as vLLM and SGLang, ensuring stable behavior under production inference patterns (prefill/decode, batching, KV cache behavior, scheduling, etc.).
• Hardware/software co-design and optimization
• Provide technical direction for HW/SW co-optimization of existing and evolving AI architectures to: maximize computational efficiency, increase sustained throughput, reduce latency and variance, improve scalability across cores, memory hierarchies, and system configurations, raise the ceiling on what Ampere platforms can deliver
• Build and evolve state-of-the-art AI accelerator software
• Contribute to and shape the architecture of software/hardware AI co-processors and accelerators, defining reusable components, reference implementations, and performance guardrails
• Cross-functional technical collaboration and influence
• Partner with compiler/runtime/kernel, platform, and systems teams to integrate and validate AI capabilities in Ampere’s platforms and accelerators from cloud to edge
• Mentorship and technical excellence
• Set engineering standards through code reviews, design reviews, benchmark methodologies, and mentorship—raising the overall bar for quality, maintainability, and performance.

About You

• Education & experience: BS Computer Science, Computer Engineering, Electrical Engineering, or Software Engineering or a related technical field & 15 years of related experience; or MS degree & 12 years; or PhD & 8 years
• Deep framework expertise: Proven experience with software development focused on PyTorch, ONNX, and llama.cpp, including integration, graph/operator enablement, and performance-focused engineering
• Linux accelerator runtime / driver experience (preferred): Experience building or extending user-mode drivers and/or runtime libraries for GPUs or deep learning accelerators on Linux is a plus
• Strong systems programming + performance tuning: Deep expertise in Python and C/C++, with a strong track record in performance engineering (profiling, optimization, throughput/latency analysis, memory behavior)
• Solid ML/AI fundamentals: Strong understanding of AI/ML concepts (neural networks, data processing frameworks), and familiarity with modern model families including Transformers and Diffusion architectures
• Fluent with modern AI development tools (preferred): Comfortable using modern AI programming tools and workflows such as Codex or Claude Code.