CUDA Kernel Engineer

Pragmatike•New York, NY

12d•Onsite

About The Position

Pragmatike is hiring on behalf of a fast-growing AI startup recognized as a Top 10 GenAI company by GTM Capital , founded by MIT CSAIL researchers. We are searching for a CUDA Kernel Engineer who has hands-on experience developing and optimizing NVIDIA CUDA kernels from scratch . You will work on the GPU performance layer powering large-scale, high-throughput AI systems used by Fortune 500 customers. This role is ideal for someone who deeply understands NVIDIA GPU architecture, memory hierarchy, warp-level execution, and profiling workflows not someone coming from generic hardware, FPGA, or non-NVIDIA compute backgrounds. You will directly influence the GPU efficiency, throughput, and scalability of mission-critical AI systems.

Requirements

Proven track record building NVIDIA CUDA kernels from scratchnot just calling existing libraries.
Strong ability to optimize kernels (tiling strategies, occupancy tuning, shared memory design, warp scheduling).
Deep understanding of CUDA threads, warps, blocks, and grids, GPU memory hierarchy and memory coalescing, as well as warp divergence (how to detect, analyze, and mitigate it)
Experience diagnosing PCIe bottlenecks and optimizing host-device transfers (pinned memory, streams, batching, overlap).
Familiarity with C++, CUDA runtime APIs, and GPU debugging/profiling tooling.

Nice To Haves

Experience with multi-GPU or distributed GPU systems (NCCL, NVLink, MIG).
Background in GPU acceleration for ML frameworks or HPC workloads.
Knowledge of model inference optimization (TensorRT, CUDA Graphs, CUTLASS).
Exposure to compiler-level optimization or PTX/SASS analysis.
Startup experience or comfort working in fast-moving, ambiguous environments.

Responsibilities

Design, implement, and optimize custom CUDA kernels for NVIDIA GPUs , with a focus on maximizing occupancy, memory throughput, and warp efficiency.
Profile GPU workloads using tools such as N sight Compute, Nsight Systems, nvprof, and CUDA‐MEMCHECK .
Analyze and eliminate performance bottlenecks including warp divergence, uncoalesced memory access, register pressure, and PCIe transfer overhead.
Improve GPU memory pipelines (global, shared, L2, texture memory) and ensure proper memory coalescing.
Collaborate closely with AI systems, model acceleration, and backend distributed systems teams.
Contribute to GPU architecture decisions, kernel libraries, and internal performance-engineering best practices.