ML Performance Engineer: Algorithm Mapping for AI Accelerators

Cerebras; United States; Remote; USD 140,000 - 220,000; Full time

Job summary

Cerebras Systems builds the world's largest AI chip and leads in fast ML training and inference. We seek an engineer to map advanced algorithms to Cerebras architecture, benchmark against GPUs, and understand trade-offs as models scale.

You will develop performance models, run experiments, and help push frontier-grade efficiency across our systems. You will combine analytical modeling with hands-on prototyping, spanning kernel-level to end-to-end performance, while evaluating new techniques and

Qualifications

• Experience building analytical and empirical performance models.
• Experience in ML training/inference workloads and performance analysis.
• Ability to design and run benchmarks across large-scale models.

Responsibilities

• Build analytical and empirical performance models for state-of-the-art ML training and inference algorithms.
• Characterize asymptotic behavior and how trade-offs scale with model size, sequence length, batch size, parallelism, and hardware.
• Construct Pareto frontiers across model quality, latency, throughput, memory, communication, and compute cost.
• Develop prototype implementations and benchmarks for Cerebras WSE and relevant GPU/software baselines.
• Analyze system behavior to identify kernel, compiler, runtime, communication, and algorithmic bottlenecks.
• Evaluate emerging techniques in parallel token processing and related areas.

Skills

Performance modeling

Benchmarking

Prototyping

Kernel-level optimization

Tools

GPU baselines

Job description

Cerebras Systems builds the world's largest AI chip and leads in fast ML training and inference. We seek an engineer to map advanced algorithms to Cerebras architecture, benchmark against GPUs, and understand trade-offs as models scale.

You will develop performance models, run experiments, and help push frontier-grade efficiency across our systems. You will combine analytical modeling with hands-on prototyping, spanning kernel-level to end-to-end performance, while evaluating new techniques and