Member of Technical Staff – GPU Performance

arrow

/ $200000 - $450000 annum

INFO

Salary

SALARY:

$200000 - $450000

Location

LOCATION

Job Type
JOB TYPE

Permanent

Research Engineer, GPU Performance (Member of Technical Staff) | Fully Remote (within 6 hours of New York time) | Up to $370,000 + equity

Make frontier world models train faster and run in real time. Your work will directly decide what is computationally possible for one of the leading generative video labs.

The Company
A late-stage AI research company building multimodal world models for video and interactive media. Its latest video generation model outperforms systems from far larger labs. The focus is now moving from static generation to real-time, interactive simulation, with applications ranging from creative tools to robotics.

The Role

  • Optimise training throughput across large GPU clusters, improving MFU through custom kernels, FP8/BF16 mixed precision, memory-efficient attention and activation checkpointing.
  • Design and maintain distributed training infrastructure, including tensor and context parallelism, FSDP and fault-tolerant multi-node setups.
  • Profile and accelerate real-time multimodal inference using CUDA graphs, KV cache optimisation, operator fusion and latency reduction.
  • Improve the efficiency and reliability of training infrastructure at scale.
  • Contribute across the stack, from low-level kernels to high-level model design.

What We're Looking For

  • 5+ years in ML systems, GPU optimisation or performance engineering. High-trajectory candidates with less experience are also considered.
  • Hands-on kernel development in CUDA or Triton.
  • Training at scale on thousands of GPUs, or demonstrable efficiency gains from constrained compute in a startup setting.
  • Deep knowledge of PyTorch or JAX internals and low-precision techniques (FP8, INT8), plus distributed systems experience with NCCL and model parallelism.
  • A hands-on individual contributor with a track record of sustained impact in previous roles.
  • Nice to have: torch.compile, inference optimisation with Flash Attention or TensorRT, generative or diffusion model infrastructure, open-source contributions and kernel competitions.

What's On Offer

  • Up to $370,000 plus ISO equity.
  • Fully remote, anywhere within six hours of New York time.
  • Visa sponsorship available.
  • End-to-end ownership of performance across the stack, without big-tech bureaucracy.
  • A low-ego, research-led team that ships quickly.

How to Apply
Apply via the link or contact Tom Brammer at Harnham.

SIMILAR
JOB RESULTS

4k-Harnham_DA copy
CAN’T FIND THE RIGHT OPPORTUNITY?

STILL
LOOKING?

If you can’t see what you’re looking for right now, send us your CV anyway – we’re always getting fresh new roles through the door.