Research Engineer, GPU Performance

Harnham; California; On-site; USD 180,000 - 240,000; Full time

Job summary

Harnham collaborates with a well-funded AI research client in California to advance multimodal models. This highly technical role focuses on making large-scale AI systems faster, more efficient, and capable of real-time operation, spanning from GPU kernels to distributed training.

You will optimize throughput, implement FP8/BF16, memory-efficient attention, and checkpointing, profile inference pipelines, and push latency down with CUDA graphing and operator fusion, shaping the future of

Qualifications

• 4+ years of experience in ML systems, infrastructure, or performance optimization.
• Strong GPU programming experience with modern frameworks.
• Distributed training experience across large clusters.
• Familiarity with PyTorch or JAX internals.

Responsibilities

• Optimize training throughput across large GPU clusters.
• Implement mixed precision (FP8, BF16), memory-efficient attention, and checkpointing.
• Design and scale distributed training systems (tensor parallelism, FSDP, multi-node).
• Profile and optimize inference pipelines for real-time multimodal generation.
• Improve latency via CUDA graphing, KV cache optimization, and operator fusion.

Skills

CUDA

Triton

GPU programming

Distributed training

PyTorch/JAX internals

Mixed precision (FP8/BF16)

Education

Bachelors or Masters in CS/EE

Tools

CUDA toolkit

PyTorch

JAX

NCCL

Job description

We’re partnered with a well-funded AI research company focused on building next-generation multimodal models for media and interactive experiences. Their work spans cutting-edge generative systems and is increasingly moving toward real-time, interactive environments, pushing beyond static outputs into dynamic, AI-driven applications.

This is a deeply technical, high-impact role focused on making large-scale AI systems faster, more efficient, and capable of running in real time. You’ll work across the stack, from low-level GPU kernels to distributed training systems, directly influencing what is computationally possible for next-generation AI models.

What You’ll Do

• Optimize training throughput across large GPU clusters, improving efficiency and utilization
• Implement techniques such as mixed precision (FP8, BF16), memory-efficient attention, and checkpointing
• Design and scale distributed training systems (tensor parallelism, FSDP, multi-node setups)
• Profile and optimize inference pipelines for real-time multimodal generation
• Improve latency through CUDA graphing, KV cache optimization, and operator fusion
• Contribute across the stack, from kernel-level optimization to system-level architecture

Requirements

• 4+ years of experience in systems engineering, ML infrastructure, or performance optimization
• Strong experience with GPU programming (CUDA, Triton, or similar)
• Experience with distributed systems and large-scale training (NCCL, model parallelism)
• Familiarity with ML framework internals such as PyTorch or JAX
• Experience with mixed or low-precision techniques (FP8, INT8, BF16)
• Proven experience building and operating scalable, fault-tolerant training systems
• Strong interest in pushing the limits of performance for cutting-edge AI systems

Nice to Have

• Experience with compiler optimizations or model compilation (e.g., PyTorch compile)
• Background working on large multimodal or generative models
• Exposure to real-time inference systems

If you're interested in working on the systems that enable next-generation AI models to train faster and run in real time, this is a rare opportunity to operate at the cutting edge of research and infrastructure.