Runtime Engineer
Posted Updated
- Normal’s hardware runs the most memory-bound parts of large-model inference inside memory itself, rather than shuttling weights to a processor
- The software that makes this programmable doesn’t exist yet in any standard form
- The runtime is the layer the rest of that software runs through: it is how compiled programs get loaded, scheduled, and executed on our hardware
- You will own its design and much of its implementation, and set its technical direction
- This is a senior, hands-on role
- Our chip is pre-silicon, so you will build the runtime against our simulators and FPGA prototypes first, behind one interface, so that software above the runtime doesn’t change when silicon arrives
- Runtime Architecture: Own the design of the runtime: device management, memory allocation and host-device transfer, executable loading, and command submission. Success means the compiler and serving layers program our hardware through one stable interface
- Hardware Abstraction: Hide the differences between simulators, FPGA prototypes, silicon, and later hardware generations behind that interface
- Device State: Manage the state that lives on our hardware between requests, such as resident model weights, so the serving layer can rely on it
- Performance: Keep runtime overhead off the critical path for LLM inference, measured by end-to-end latency and throughput
- Cross-Team Interfaces: Define the boundary with the compiler above the runtime, and with the driver and firmware below it, which our hardware team owns
- Technical Leadership: Set the runtime’s technical direction, review designs that depend on it, and work with compiler and serving engineers so the interface serves them
- Experience building runtimes, drivers, or hardware abstraction layers for accelerators or other non-standard hardware
- Understanding of LLM inference workloads and what they demand of a runtime
- Track record of owning a complex systems component end to end, from design through production, and driving technical decisions across teams
- Strong systems programming in C++ or Rust: memory management, concurrency, and low-level performance
- Comfort building software for hardware that is still evolving, from simulation through bring-up
- Contributions to open-source runtime or systems infrastructure
- Experience integrating a runtime with an inference server such as vLLM
- Experience with in-memory compute or other non-von-Neumann hardware
- Experience with open runtime designs such as IREE’s HAL, or with CUDA or ROCm runtime internals