Software Engineer, High Performance Computing
$150,000–$250,000 year
On-siteSan Francisco, California, United States
Job Summary
Design and build the video-native dataloader that returns tensors directly to the GPU at line rate, featuring rank-aware sampling, NVMe caching, and decode pipelining. Profile and optimize the full data path from object store through device RAM, eliminating avoidable copies and stalls to saturate B200, GB200, and NVL72 hardware. Own performance benchmarks against customer baselines and historical numbers, catching regressions at PR time while partnering with researchers to land the loader in their training stacks. Work cross-team with Storage Infrastructure and Visual Understanding teams to define index boundaries and model-output ingestion paths. Join the small team in the SF Mission district office, 4 days/week, where we train Physical AI systems on petabytes of multimodal data.
Required Qualifications
- Obsession with systems-level performance
- Strong opinions on io_uring
- Live and breathe Rust, C++, or C
- Strong familiarity with operating systems
- A sense for where bytes actually go
Desired Qualifications
- Experience working with GPUs
- Experience working with SLURM, Kubernetes for GPU workloads, or other HPC schedulers
- Hands-on CUDA experience
- Deep expertise on memory and caching subsystems
- Worked on video decode pipelines (PyAV, decord, NVDEC) or PyTorch DataLoader internals
- Contributed to open-source systems projects in Rust/C++
Hiring someone like this?
Get your role in front of qualified candidates on Sorce.