Research Engineer, Normal Computing

• Deep understanding of large model inference: attention mechanisms, KV cache, long-context decoding, memory bandwidth constraints
• Experience with inference optimization: quantization, sparsity, kernel fusion, or memory-efficient attention
• Familiarity with stochastic systems, probabilistic methods, numerical analysis, or analog computation
• Experience implementing algorithms close to hardware, not just in high-level frameworks
• Comfort reasoning from first principles about what a novel substrate can do efficiently
• Track record of taking ideas from theory to working implementation on real hardware
• Strong programming skills in Python and at least one systems language
• Collaborative instinct and ability to work across hardware, architecture, and software teams

Desirable

• You will develop the computational methods that make AI inference run efficiently on Normal's thermodynamic hardware
• The core challenge is not adapting standard GPU kernels to a new chip
• It is rethinking how operations like attention, memory access, and long-context decoding behave when the underlying substrate uses stochastic analog computation in memory rather than conventional digital logic
• Normal's ASICs run the heaviest operations of large model inference inside memory itself
• Your job is to develop the algorithms that exploit this natively: understand what transformer and diffusion workloads are well-suited to stochastic analog execution, design numerical methods that map onto the hardware's physical dynamics, and validate them against real silicon or high-fidelity simulation
• This is a co-design role
• The hardware and the algorithms are developed in parallel, which means you will influence architectural decisions, not just implement against a fixed specification
• The strongest candidates have a deep understanding of both large model inference and the mathematics of stochastic systems, and have built systems that run on real hardware, not just in theory
• Algorithm Development: Develop algorithms for transformer inference workloads running on stochastic analog processing-with-memory hardware
• Software/Hardware Co-Design: Work directly with hardware and architecture teams to shape what the chip can and should compute natively
• Numerical Methods: Design numerical methods that exploit thermal noise and analog dynamics rather than working around them
• Evaluation & Benchmarks: Build evaluation frameworks and benchmarks that characterize algorithm behavior on real hardware or simulation
• Workload Translation: Translate insights about model workloads into constraints and opportunities for hardware design
• Rapid Prototyping: Prototype and iterate rapidly as hardware evolves from simulation to silicon