The role asks for someone who can find out why large AI training runs on GPU clusters waste compute, and then fix it. That means profiling jobs, tuning GPU kernels and communication between machines, and building dashboards that show how well the hardware is used. A strong candidate has a few years of hands-on performance or HPC work, knows CUDA or Triton well, and is comfortable working alone at a young company.
How it fits the AI era
AI models are the thing being optimized here, so the job exists because of the AI era. Techniques like quantization, pruning and distillation are part of the daily toolkit, and AI coding tools may speed up routine profiling and scripting. Judgment about trade-offs, such as keeping scientific accuracy while changing a model to run faster, and working out what the hardware is really doing, stays with the person.
How it builds AI experience
The work gives direct experience with large training and inference for chemistry and physics models, which is a less common area than language models. Someone would learn how scaling behaves on big GPU clusters and how model design choices affect speed and cost. Working beside researchers who build the materials search engine also shows how efficiency work turns into faster scientific testing.
Worth showing
- CUDA or Triton kernel tuning
- Distributed training at scale
- MFU improvement results
- Open-source performance contributions
This is our own read, written from the employer’s posting and the company’s Spotlight. The employer’s posting is the source of truth for the role.
About CuspAI
An AI search engine for new materials that the company says turns a wanted set of physical properties into synthesizable chemical designs in months instead of years of lab trial and error.
#J-18808-Ljbffr…