About the role
We're building a new class of Scientific Processing Units (SPUs) to push the boundaries of High-Performance Computing (HPC). In this role, you'll develop and optimize the core computational kernels needed to run scientific applications across a wide range of domains on our custom parallel SPU architecture, from simulation and emulation to real hardware.
You'll work across abstraction layers. Starting from the mathematics and algorithms behind an application, building high-level reference implementations, translating them into parallel kernels, and optimizing them against our architecture. You should be comfortable moving between a mathematical description of a problem, C++/Python reference code, low-level parallel software, and the hardware that executes it.
This is not an easy role. It requires passion for understanding how hardware and software work together, the ability to wear multiple hats, strong communication skills, a lot of patience, independence, and, most importantly, zero ego. Why? Because this is the first time something like this has ever been attempted.
Responsibilities
-
Develop high-level reference implementations of scientific applications (e.g. finite-difference time-domain methods, computational fluid dynamics, electromagnetic wave propagation, etc.) from mathematical and algorithmic descriptions.
-
Implement and parallelize scientific applications using our proprietary Software Development Kit (SDK) across SPU simulation, emulation, and real-hardware environments.
-
Develop and optimize performance critical kernels for the SPU architecture.
-
Own iterative performance optimization loops, benchmark, profile, identify bottlenecks, implement improvements and repeat, until we reach our performance targets.
-
Understand applications across abstraction layers, from their mathematical formulation to their mapping onto parallel hardware.
