AI Systems, Model Optimization

Unconventional, Inc.•Palo Alto, CA

2d•Onsite

About The Position

Since 2022, AI has entered the mainstream, reshaping entire industries from education and software development to fundamental consumer behaviors. This revolution has created an unprecedented demand for computation - a demand that is now fundamentally limited by energy, not just in the datacenter, but at a global scale. At Unconventional, our mission is to solve this. We are rethinking computing from the ground up to build a new foundation for AI that is 1000x more efficient. We're doing this by exploiting the rich physics of semiconductors, mapping neural networks directly to the device physics rather than relying on layers of inefficient abstraction.

Requirements

An MS/PhD or equivalent research/project experience in a quantitative field such as AI/Machine Learning, Computer Science, Physics, Electrical Engineering, or Applied Math.
Deep, practical understanding of the modern AI/ML stack and optimized compilation and execution of algorithms on modern GPU systems.
Proven experience in profiling, identifying, and resolving performance bottlenecks in complex ML codebases.
Demonstrated ability to map state-of-the-art AI model architectures (e.g., Transformers, Mixture of Experts, diffusion models) to system performance implications and apply advanced efficiency techniques such as sparsity, quantization, and distillation.
Deep experience with PyTorch, including its internals, torch.compile, and distributed data parallel (DDP) / fully sharded data parallel (FSDP) libraries.

Nice To Haves

Experience with production-grade training frameworks (e.g., Megatron-LM, DeepSpeed) and distributed training at scale.
A forward-looking perspective on co-designing training systems for unconventional computing paradigms that map closely to the physics of underlying systems.
Research or practical experience in advanced approximation/compression techniques beyond standard quantization, including noise-aware or physics-constrained training.

Responsibilities

Develop rigorous performance models to evaluate compute, memory, and energy trade-offs.
Track pareto-optimality across models and hardware configurations.
Drive the partitioning and mapping of complex AI models down to hardware.
Apply and invent advanced optimization strategies from first principles, including custom quantization schemes, sparsity/pruning, and distillation to fit the physical constraints of our substrates.
Develop and apply Quantization-Aware Training (QAT), noise-aware training, and sparsification techniques to adapt models to the physical constraints of our analog compute substrates, including memory footprint, connectivity, precision, and noise.
Develop and optimize kernels using low-level programming models like CUDA, Triton, or CUTLASS.
Profile and debug complex ML codebases to resolve performance bottlenecks (training and inference).
Act as a translator between AI model architects and hardware/infrastructure engineering teams, converting model requirements into concrete specifications and codifying learnings for tapeouts.