Research Engineer - CUDA Kernel Engineering

Job Title

Research Engineer — CUDA Kernel Engineering

Role Summary

Develop and optimize CUDA kernels and integration layers to enable large-scale training, inference, and reinforcement-learning systems for AI-driven semiconductor design and verification. Work with researchers and engineers to maximize GPU efficiency across multi‑GPU deployments and deliver production-ready kernels, benchmarks, and tooling.

Expect to build performance benchmarks, CI/test suites, and contribute selected kernels and tools to open-source AI/HPC ecosystems.

Experience Level

Mid-level.

Responsibilities

Core responsibilities include:

• Design, implement, and optimize CUDA kernels for compute- and memory-bound primitives (attention, routing, graph operations, physics-inspired operators, etc.).
• Profile GPU workloads and resolve performance bottlenecks (compute, memory, interconnect, occupancy).
• Integrate custom kernels and operators into training and inference frameworks (e.g., PyTorch, Megatron, vLLM, TorchTitan).
• Develop multi‑GPU scaling strategies and handle communication (NVLink, NCCL) and memory management for large models.
• Build benchmarks, tests, and automation to validate correctness and measure performance across hardware generations.
• Collaborate with AI researchers and semiconductor domain experts to translate domain-specific algorithms into high-performance GPU code.
• Package, document, and maintain kernels and tooling for internal use and selected open-source releases.

Requirements

Key qualifications for the role. Separate must-have and nice-to-have items are listed below.

Must-have

• Practical experience writing and optimizing CUDA kernels for large-scale AI workloads.
• Strong skills in C++ and CUDA, and experience with GPU performance tools and workflows (Nsight, CUPTI, nvprof, profilers).
• Experience integrating custom CUDA ops into PyTorch or comparable training/inference frameworks.
• Familiarity with NVIDIA hardware and software stack (recent architectures, NVLink, NCCL, Triton).
• Experience with multi‑GPU scaling and inter-GPU communication patterns.
• Ability to create benchmarks, tests, and CI for kernel correctness and performance.

Nice-to-have

• Experience with Triton or other GPU code-generation frameworks.
• Background in graph reasoning, symbolic computation, or hardware simulation kernels.
• Previous contributions to open-source AI or HPC projects.
• Familiarity with reinforcement learning training systems or large-model engineering.

Education Requirements

Not specified.

About the Company

Company: Voltai

Headquarters: Palo Alto, CA, United States

Voltai develops AI-driven world models and agents to design, evaluate, and optimize physical systems—focusing on hardware, electronics, and semiconductors to enable AI-led hardware co-design, performance modeling, and cross-domain optimization.