Teams at Neurophos

Model the architecture and performance of an optical AI inference accelerator through hardware/software co-design. Responsibilities include bringing up transformer and other AI workloads, integrating Hugging Face and PyTorch models, developing Python and C++ functional, performance, energy, and power models, conducting roofline and limiter analyses, modeling memory and compute blocks, simulating RTL, and maintaining reproducible tests, configurations, plots, and reports.

Build functional, performance, energy, power, and area models for AI accelerator workloads and hardware/software co-design. Bring up PyTorch and Hugging Face workloads, model optical GEMM, memory systems, tiling, scheduling, ISA, NoC traffic, and multi-chip mapping. Develop Python analytical models and bit-accurate C++ simulations, contribute to event-driven simulation infrastructure, correlate models with RTL through Verilator and SystemVerilog, establish modeling methodology, and mentor engineers.

Own performance and energy benchmarking for an optical AI inference accelerator. Build reproducible benchmarks across analytical models, architecture models, RTL simulation, and competitor GPUs. Bring up workloads from Hugging Face, PyTorch, papers, and inference stacks; measure latency, throughput, power, and energy; analyze bottlenecks; operate cloud or lab environments; and document configurations, logs, assumptions, and discrepancies.