The company is building a new kind of Scientific Processing Unit (SPU) to speed up high-performance computing. The role focuses on the core computational kernels that run scientific applications across many fields on the company's custom parallel SPU architecture, in simulation, emulation, and on real hardware.
You will work across layers of the stack. You start from the math and algorithms behind an application, write high-level reference implementations, turn them into parallel kernels, and tune those kernels for the architecture. You need to be at home with mathematical problem descriptions, C++ and Python reference code, low-level parallel software, and the hardware that runs it.
The work is demanding. It calls for curiosity about how hardware and software fit together, the willingness to take on many roles at once, clear communication, patience, independence, and above all a willingness to set ego aside, since this has not been tried before.
Responsibilities
Write high-level reference implementations of scientific applications such as finite-difference time-domain, computational fluid dynamics, and electromagnetic wave propagation, starting from mathematical and algorithmic descriptions.
Parallelize and implement scientific applications with the company's proprietary Software Development Kit (SDK) across SPU simulation, emulation, and real-hardware environments.
Build and tune performance-critical kernels for the SPU architecture.
Run repeated optimization cycles: benchmark, profile, find bottlenecks, make changes, and repeat until the performance targets are met.
Follow applications from their mathematical form down to how they map onto parallel hardware.
Find ways to improve the compilation flow, runtime, hardware utilization, and overall system efficiency.
Work with architecture and performance teams to analyze bottlenecks and design software and hardware solutions together.
Identify problems and opportunities on your own, propose next steps, and move work forward without step-by-step direction.
Write clear, concise technical documentation for the company's engineers and customers.
Keep up with parallel programming models, HPC architectures, numerical methods, and ways of mapping scientific workloads onto parallel hardware.
Requirements
Bachelor's degree in Computer Science, Electrical Engineering, Applied Mathematics, Physics, or a related field.
Master's or PhD in Computer Science, Electrical Engineering, Applied Mathematics, Physics, or a related field.
5+ years of modern C++ experience for parallel programming and HPC. CUDA C++ experience is strongly preferred. Strong candidates with fewer years are also welcome.
Able to read numerical algorithms and turn mathematical descriptions into working code. You do not need to be a mathematician, but gradients, divergence, stencils, and multidimensional discretization should be familiar territory.
Able and willing to debug across layers, from the application or numerical algorithm through the software and down into the architecture.
Solid grasp of computer architecture and how software interacts with hardware.
Proficient in C++, Python, and Linux development environments.
Excellent written and verbal communication.
Works well alone and in a team, takes ownership of unclear problems, and decides what to do next.
Willing to do the hard work needed to bring the SPU to life.