Research Engineer: Algorithms for Stochastic AI Hardware

Normal Computing; New York (NY); On-site; USD 150,000 - 210,000; Full time

Job summary

Normal Computing is developing silicon that uses thermal noise as a computational resource to accelerate AI inference. You will design algorithms and numerical methods that map to hardware's analog dynamics, enabling efficient transformer and diffusion workloads.

The role is a co-design effort where software, hardware, and architecture teams collaborate in parallel, influencing architectural decisions rather than following a fixed spec.

Qualifications

• Deep understanding of large model inference, including attention and long-context decoding.
• Experience with inference optimization, such as quantization or memory-efficient attention.
• Familiarity with stochastic systems or analog computation.
• Experience implementing algorithms close to hardware.

Responsibilities

• Develop algorithms for transformer inference on stochastic analog hardware.
• Work with hardware and architecture teams to shape native computation.
• Design numerical methods that exploit analog dynamics.
• Build evaluation frameworks and benchmarks on real hardware or simulation.
• Translate model workload insights into hardware design constraints.
• Prototype and iterate as hardware evolves from simulation to silicon.

Skills

Attention mechanisms

KV cache

Memory bandwidth

Inference optimization

Analog computation

Python

Systems programming

Collaboration

Tools

None

Job description

Normal Computing is developing silicon that uses thermal noise as a computational resource to accelerate AI inference. You will design algorithms and numerical methods that map to hardware's analog dynamics, enabling efficient transformer and diffusion workloads.

The role is a co-design effort where software, hardware, and architecture teams collaborate in parallel, influencing architectural decisions rather than following a fixed spec.