# Driver Engineer at Modular

- Company: Modular
- What the company does: The unified AI inference stack - from custom GPU kernels to production cloud serving on NVIDIA and AMD. 2x performance. Top open models. Open source stack. Backed by General Catalyst, Greylock and GV.
- Company website: https://www.modular.com/
- Type: Startups
- Level: Mid level
- Location: United States - Remote
- Work setup: Remote
- Posted: 2026-05-28
- Apply by: 2026-10-08
- Apply: https://jobs.gem.com/modular/am9icG9zdDqfHQlU8GiJliF30Lu6Ap_l
- Page: https://www.1752.vc/careers/jobs/modular-driver-engineer/

## About the role

As a Driver Engineer, you'll work on the layer that sits between Modular's compiler and runtime stack and the silicon — the devices, contexts, queues, synchronization primitives, memory allocators, networking, and collectives that every kernel and graph execution depends on. Your work directly affects how reliably and efficiently MAX and Mojo run across NVIDIA, AMD, Apple Silicon, and emerging accelerators, and how cleanly kernel and graph authors can target them from Mojo and Python.

## What they're looking for

- 3+ years of experience writing high-performance, low-latency production systems C++ (modern C++17/20), with strong instincts for ownership, lifetime, ABI stability, concurrency, and parallelism
- Working understanding of accelerator execution and memory models — stream ordering, host↔device transfers, and pinned memory. Exposure to peer access and multi-GPU topologies (NVLink/xGMI, NUMA) is a plus
- Solid instincts for library and API design — naming, layering, ergonomics
- Strong debugging skills in asynchronous, multi-device systems — GDB/LLDB, sanitizers, CUDA/HIP error modes, race conditions, leaked resources
- A proactive, collaborative mindset, intellectual curiosity, and drive to identify and solve complex technical challenges as part of a high-performing team
- Experience with asynchronous runtimes, custom memory allocators, or RDMA-based networking (Infiniband, RoCE, EFA, UCX, NIXL, MPI, NVSHMEM/ROCSHMEM/OpenSHMEM)