Senior Modeling Architect, Benchmarking

About Neurophos

The demand for new data centers and AI compute is rapidly outpacing the planet's energy capacity. Digital solutions are hitting a power wall as we approach the physical limits of traditional silicon. Conquering this bottleneck means rethinking the fundamental architecture of inference compute. The industry's current path can't meet the need, so we're taking a different approach.

Instead of traditional electronic circuits, we use silicon photonics and an active, programmable metasurface to perform matrix multiplications at the speed of light. Our optical cells are 10,000x smaller than traditional photonic components, enabling unprecedented density. By using photonics instead of electricity, our chips become more efficient as they scale. This architecture will deliver up to 100 times the energy efficiency of existing solutions while significantly improving performance for large-scale AI inference.

We’ve assembled a world-class team of industry veterans and recently raised a $110M Series A led by Gates Frontier. Participants include M12 (Microsoft’s Venture Fund), Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others.

Join us and shape the future of computing!

Location: Austin, TX or Sunnyvale, CA. Full-time onsite position.

Reports To
:
Sr. Director of Modeling

FLSA Status
:
Exempt

We are seeking a performance engineer to own the benchmarking numbers behind the T100 optical inference accelerator. Architecture and product decisions here are made on measured performance and energy, and this role produces those figures for the same workloads at every level of fidelity we use: roofline and limiter analysis, architecture performance models, in-house RTL simulation, and measured runs on competing GPUs and accelerators.

You will join the Architecture and Modeling team, set the measurement methodology, and keep it current as models, software stacks, drivers, and hardware generations turn over. Every result ships with the harness, config, plots, logs, and assumptions behind it, so anyone can rerun it and see how the number was reached.

• Own the performance and energy metrics that architecture, product, and leadership rely on, and keep them consistent across modeling fidelities and measured hardware.

• Produce numbers for the same workloads across all fidelities we use internally: roofline and limiter models, architecture performance models, in-house RTL simulation, and measured competitor hardware. Work with the modeling team to keep the simulated and measured workload sets aligned.

• Hold workload definitions constant across fidelities, including model or application, sequence length, batch, precision, prefill versus decode, and tensor, pipeline, and sequence parallelism.

• Bring up inference workloads from Hugging Face, PyTorch, published papers, and vendor stacks such as vLLM, SGLang, TensorRT-LLM, and Triton Inference Server. These include dense and Mixture-of-Experts (MoE) transformers, attention and KV-cache reduction strategies, and hybrid/SSM models, as well as retrieval, speech, vision, and recommendation workloads that map onto the accelerator.

• Measure competing GPUs and accelerators end-to-end, owning the cloud or lab account, image, drivers, and run recipe.

• Report time to first token (TTFT), inter-token latency (ITL), tokens per second, tokens per second per watt, and energy, using nvidia-smi, DCGM, power capping, or equivalent instrumentation.

• Document where RTL simulation, the performance model, and measured competitor results disagree, and attach the configs and logs behind each.

• Maintain a reviewed internal benchmark suite. Keep internal-only results clearly separate from anything cleared for customer…