Staff Software Engineer - GenAI Performance and Kernel

P-1285

About This Role

As a staff software engineer for GenAI Performance and Kernel, you will own the design, implementation, optimization, and correctness of the high-performance GPU kernels powering our GenAI inference stack. You will lead development of highly-tuned, low-level compute paths, manage trade-offs between hardware efficiency and generality, and mentor others in kernel-level performance engineering. You will work closely with ML researchers, systems engineers, and product teams to push the state-of-the-art in inference performance at scale.

What You Will Do

g. attention, MLP, softmax, layernorm, memory management) optimized for various hardware backends (GPU, accelerators) Drive the performance roadmap for kernel-level improvements: vectorization, tensorization, tiling, fusion, mixed precision, sparsity, quantization, memory reuse, scheduling, auto-tuning, etc. g.

g. memory layout, dataflow scheduling, kernel fusion boundaries) Mentor and guide other engineers working on lower-level performance, provide code reviews, help set best practices Collaborate with infrastructure, tooling, and ML teams to roll out kernel-level optimizations into production, and monitor their impact What We Look For BS/MS/PhD in Computer Science, or a related field Deep hands-on experience writing and tuning compute kernels (CUDA, Triton, OpenCL, LLVM IR, assembly or similar sort) for ML workloads Strong knowledge of GPU/accelerator architecture: warp structure, memory hierarchy (global, shared, register, L1/L2 caches), tensor cores, scheduling, SM occupancy, etc. g.

Compensation

practices. The pay range(s) for this role is listed below and represents the expected salary range for non-commissionable roles or on-target earnings for commissionable roles.

packages are based on several factors that are unique to each candidate, including but not limited to job-related skills, depth of experience, relevant certifications and training, and specific work location. Based on the factors above, Databricks anticipates utilizing the full width of the range.

package for this position may also in