Principal AI Systems Performance Engineer
On-siteSan Jose, California, United States
Job Summary
Bring up and optimize cutting-edge foundation models on the SambaNova platform through the full software stack. Profile and enhance model performance across compiler, runtime, and hardware layers to achieve state-of-the-art throughput and latency. Collaborate with machine learning, compiler, runtime, and hardware teams to deliver co-designed, high-performance AI applications. Integrate latest advances in model architecture, quantization, scheduling, and memory optimization from both academia and industry. Develop robust, scalable, and efficient end-to-end inference solutions aligned with customer needs. Identify performance bottlenecks and propose dataflow or scheduling optimizations for both single-node and distributed systems. Work hands-on with frontier architectures like DeepSeek R1 and GPT OSS to push the limits of large-scale AI inference.
Required Qualifications
- Bachelor's or higher degree in computer science, electrical engineering, or a related field (e.g., applied mathematics, physics, or statistics)
- 3+ years of experience in one or more of the following areas: Deep learning model development and performance optimization, Compiler, runtime, or kernel-level optimization, Software–hardware co-design or systems performance tuning
- Proficiency in Python or C++, with strong foundations in algorithms, data structures, and numerical computing
- Experience with at least one major ML framework — PyTorch, TensorFlow, or JAX
- Demonstrated ability to analyze and optimize performance in real-world ML pipelines
Desired Qualifications
- Hands-on experience with LLM or multimodal model training and inference
- Background in large-scale distributed training, continuous batching, and high-throughput inference systems
- Familiarity with quantization, graph optimization, kernel fusion, and model partitioning
- Experience with frameworks such as DeepSpeed, Megatron, vLLM, or TensorRT
- Strong GPU programming skills (CUDA, Triton, or OpenCL); experience with cuDNN, cuBLAS, or similar libraries is a plus
- Knowledge of memory hierarchy optimization, caching, and scheduling for large-scale model execution
- Publication record or open-source contributions in ML systems or performance optimization is a plus
Hiring someone like this?
Get your role in front of qualified candidates on Sorce.