GPU Performance Engineer | Experienced Hire

Susquehanna International Group; Bala Cynwyd (PA); On-site; USD 100,000 - 130,000; Full time

Job summary

Susquehanna International Group is seeking a GPU Performance Engineer in Bala Cynwyd, Pennsylvania. In this role, you will design and optimize CUDA kernels for low-latency inference workloads, working closely with quantitative researchers. This position requires strong expertise in GPU architecture and programming in C/C++. A PhD in a quantitative field is preferred. The role focuses on enhancing performance for various structured inference workloads and demands a solid understanding of GPU hardware and optimization techniques.

Qualifications

• Strong proficiency in writing and optimizing CUDA kernels.
• Deep understanding of GPU architecture.
• Experience with quantitative research teams or financial models.

Responsibilities

• Design, implement, and optimize custom CUDA kernels.
• Analyze research models to identify opportunities for optimization.
• Collaborate with quantitative researchers to translate models into compute pipelines.

Skills

Proficiency in CUDA

C/C++ programming

Problem-solving skills

GPU architecture knowledge

Education

PhD in mathematics, physics, computer science, engineering, or related quantitative field

Tools

ONNX Runtime

TensorRT

Job description

Overview

We are looking for a GPU Performance Engineer to build highly optimized CUDA kernels for low‑latency inference. This role is focused on workloads where off‑the‑shelf runtimes and vendor libraries do not fully exploit the structure of the model, and where custom kernels, memory layouts, and execution strategies can deliver meaningful gains.

You will work closely with quantitative researchers and engineers to understand model structure, identify computational bottlenecks, and turn mathematical ideas into production‑grade GPU implementations. You will use your understanding of GPU hardware to help shape models that are both mathematically effective and efficient to run. The problems span compact neural networks, tree‑based models, and other structured inference workloads where latency, throughput, and efficiency all matter. This role is a strong fit for someone who enjoys low‑level optimization, performance analysis, and translating abstract models into hardware‑efficient code.

What You’ll Do

• Design, implement, and optimize custom CUDA kernels for latency‑critical inference workloads
• Develop fine‑grained GPU implementations tailored to specific model structures
• Analyze quantitative research models and computational bottlenecks to identify opportunities for parallelization and hardware‑efficient execution
• Collaborate directly with quantitative researchers to translate mathematical models into high‑performance compute pipelines
• Optimize end‑to‑end inference performance through kernel tuning, memory‑layout design, execution strategy, I/O optimization, and precision tradeoffs
• Profile and benchmark GPU performance
• Improve latency and throughput in production inference systems
• Contribute to GPU architecture decisions and performance best practices

What We’re Looking For

• Strong proficiency in writing and optimizing CUDA kernels
• Solid programming experience in C/C++ (preferred)
• Deep understanding of GPU architecture, including memory hierarchy, SIMT execution, occupancy, and latency/throughput tradeoffs
• Ability to reason about numerical stability, precision, performance tradeoffs, and how model design choices affect hardware efficiency
• Strong problem‑solving skills and comfort working with low‑level systems

Preferred Qualifications

• PhD in mathematics, physics, computer science, engineering, or related quantitative field
• Strong background in linear algebra, probability, numerical methods, or scientific computing
• Experience working with quantitative research teams or financial models
• Demonstrated ability to improve real‑world inference performance beyond baseline framework or library implementations
• Familiarity with PTX‑level behavior, tensor core utilization, or architecture‑specific tuning
• Exposure to ONNX Runtime, TensorRT, Triton, TVM, or similar systems
• Exposure to neural networks, tree‑based models (e.g., LightGBM), state space models (e.g., Mamba architectures), and experience with kernel fusion, custom operators, model compilation, or graph‑level optimization