Systems Performance Modeling Engineer

Tensordyne•Sunnyvale, CA

5d•Remote

About The Position

We are looking for a Systems Performance Modeling Engineer to build the models and tools that predict how generative AI inference workloads perform on Tensordyne systems, from a single accelerator up through rack, pod, and cluster scale. This is a hands-on engineering role for someone who likes writing simulator code, running experiments, and digging into why a prediction and a measurement don't match. Working closely with our architects and the silicon, hardware, networking, and software teams, you'll capture workload behavior, extend simulation and analytical models of our silicon, interconnect, and multi-hop fabrics, and validate them against real hardware. You'll be comfortable moving across the stack, from the model graph through collectives to the network fabric, to track down where performance is going.

Requirements

Hands-on experience building performance models, simulators, or analytical tools for ML workloads, distributed systems, or computer architecture.
Solid understanding of distributed ML execution, including parallelism strategies, collective communication (All-Reduce, All-Gather, All-to-All, etc.), and how they scale.
Working knowledge of system architecture across compute, memory, interconnect, and networking, and the ability to reason about bottlenecks between them.
Experience comparing model predictions against real measurements, and debugging where they diverge.
Strong programming skills in C++ and Python, with clean, testable, maintainable code.
Ability to take a loosely defined performance question, break it into experiments, and deliver results with minimal hand-holding.
Clear written and verbal communication, especially when presenting data and trade-offs to other engineers.
MS or higher in Computer Science, Computer Engineering, Electrical Engineering, or a related field.

Nice To Haves

Familiarity with LLM inference serving: batching, KV-cache management, disaggregated prefill/decode, and latency/throughput trade-offs.
Experience modeling or benchmarking collective communication libraries (NCCL, RCCL, or similar) on real clusters.
Background in data center or HPC networking: topologies, RDMA/RoCE, and congestion behavior.
Experience profiling ML workloads on accelerators (GPUs, TPUs, or custom ASICs).
Exposure to hardware/software co-design or early-stage architecture evaluation.
Publications or open-source contributions in ML systems, architecture, or networking.

Responsibilities

Implement and extend simulation-based performance models for multimodal generative AI inference at rack, pod, and cluster scale, covering compute, memory, collective communication, and network fabric.
Model how serving strategies (tensor, pipeline, and expert parallelism, prefill/decode disaggregation, batching, and KV-cache placement) interact with Tensordyne silicon and fabric topology, and measure the effect on latency, throughput, and cost per token.
Build trace-capture and replay tooling that records real execution from our inference runtime and replays it under hypothetical silicon, system, and network configurations.
Model collective communication on multi-hop scale-out fabrics, including implementing custom collective algorithms designed for our topology.
Run calibration experiments on Tensordyne hardware as systems come up, compare them against model predictions, and fix the sources of error.
Run design-space sweeps and write up clear analyses that architects and engineering teams use in ASIC, fabric, and system configuration decisions.
Produce performance projections that support product and customer discussions.
Keep the modeling codebase fast, tested, and reproducible so other engineers can run it themselves.