GPU Engineer (C++/CUDA)

GPU Engineer (C++/CUDA) at Scailium, Moscow. Full-time, remote/office/hybrid. Develop CUDA/HIP kernels for relational operators and tune performance on NVIDIA/AMD GPUs. Compensation discussed after interview.

Project: Scailium is building a GPU-accelerated relational database engine. Responsibilities include: - develop and maintain CUDA/HIP kernels for relational operators: JOIN, sorting, aggregations, string and window functions; - develop the execution framework — device memory pool, CUDA streams scheduling, kernel fusion, keeping intermediate columns in GPU memory between operators; - work with the compression subsystem: codecs RLE, delta-RLE, dictionary, LZ4, P4D, integration nvCOMP/hipCOMP — decompression is on the critical path of scanning; - profile and tune performance for each generation of NVIDIA accelerators (A100 → H100 → H200 → Blackwell) and AMD Instinct MI; - handle pilot deployment requests: customer query profiling, root-cause analysis, and fixes; - conduct code review of changes prepared by AI agents and bring them to merge into the main branch. The team works with AI agents: mechanical volume — mass edits across hundreds of files, test generation, tuning sweeps — is performed by agents; the engineer focuses on architectural decisions, performance tuning, and verification. Unverified code is not merged into the main branch. The share of code reading and evaluation is noticeably higher than usual.

- 5+ years developing modern C++ (C++17), including 2+ years of GPU programming in production systems; - confident CUDA: custom kernels, memory hierarchy, CUDA streams and copy/compute overlap, occupancy, register pressure; - experience with Thrust and CUB — the engine codebase is built around these libraries; - systematic profiling (Nsight Systems / Nsight Compute or rocprof) and device-level debugging (compute-sanitizer, cuda-gdb); - CMake, Linux, Docker; ability to quickly get up to speed in a large existing codebase; - technical English: code, documentation and change discussions are conducted in English.

- Full-time employment; - indefinite contract; - work format: remote, office or hybrid by agreement; - dedicated access to servers with AMD Instinct MI300X and NVIDIA accelerators from the first week; - compensation level discussed after interview.