Research Engineer

Research Engineer, AI Cluster Performance

About XPerf

XPerf builds software that validates, operates, and optimizes artificial intelligence (AI) clusters built on graphics processing units (GPUs) and other accelerators. Our products are ClusterReady for pre-deployment cluster validation, ClusterBeacon for runtime observability, and ClusterMarshal for scheduling and prediction. We were founded by ex-Intel engineers who have deployed clusters with thousands of accelerators.

The role

This is a research role. You will build a simulator that lets us benchmark the performance and energy consumption of jobs and workloads, and optimize how they use resources. ASTRA-sim is at the core of this work, and we also use higher-level simulators that model whole clusters and data centers, including job scheduling, power, and cooling.

What you'll do

• Build a simulator that models how jobs and workloads run on GPUs and other accelerators
• Model energy consumption alongside performance
• Benchmark workloads in the simulator and on real hardware, and use the measurements to keep the simulator accurate
• Use the results to optimize resource use and energy consumption
• Support technical research on our products when needed

Work environment

We are a small technical team, and we move fast. Our customers are overseas, so working hours are flexible and some work falls outside the usual US workday.

Basic requirements

• Master's degree or higher in Electrical and Computer Engineering or Computer Science
• Solid coding skills
• Basic knowledge of GPU architecture and other AI accelerators
• Understanding of the training and inference process on GPUs, and how each consumes energy
• Understanding of large language model (LLM) algorithms
• Understanding of how simulators work
• Ability to master AI tools quickly and use them in your daily work
• Ability to turn open research questions into clear requirements
• Comfort with a fast pace and flexible hours

Preferred qualifications

Education

• Doctor of Philosophy (PhD) in Electrical and Computer Engineering or Computer Science

Simulators

• Experience building or extending a simulator
• Experience with or knowledge of ASTRA-sim, our core tool for distributed training performance
• Experience with or knowledge of any related simulator, such as:
  • SustainCluster for energy- and carbon-aware AI scheduling across data centers
  • ExaDigiT for data center digital twins covering job scheduling, power, and cooling
  • Blox for cluster scheduling
  • SimAI for full-stack LLM training and inference
  • Vidur for LLM inference
  • LLMServingSim for LLM serving performance and power
  • Calculon for analytical transformer performance modeling
  • ATLAHS for network simulation from application traces
  • ttsim for Tenstorrent chips
• Experience checking simulated results against real measurements

Benchmark testing

• MLPerf Training and MLPerf Inference, including LoadGen
• GPU communication tests: nccl-tests and rccl-tests
• Network and storage tests: ib_write_bw, iperf3, and fio
• LLM serving benchmarks with vLLM or SGLang, measuring tokens per second, time to first token (TTFT), and latency
• GPU monitoring with NVIDIA Data Center GPU Manager (DCGM)

Energy

• Measuring power and energy with tools such as nvidia-smi, amd-smi, or Running Average Power Limit (RAPL)
• Modeling energy consumption in a simulator

Accelerators and resource optimization

• Hands-on experience with accelerators other than GPUs
• Job scheduling and resource allocation on clusters that mix accelerator types
• Measuring utilization and useful work completed

Coding

• 5+ years of coding experience

Equity & benefits: Meaningful stock options. Solid health/dental/vision, flexible time off, professional development support.

Direct candidates only, no recruiters.