SoC Firmware Engineering Manager, Annapurna Labs Machine Learning Acceleration, AWS

When a new Trainium or Inferentia chip comes back from the fab, our code is the first software to touch it. We're looking for a hands-on engineering manager who lives and breathes low-level software — someone who's debugged register-level issues at 3am and wants to build a team that does it better.

Our SoC HAL (Hardware Abstraction Layer) team owns the lowest layer of user-space software on AWS's custom ML accelerator chips: the firmware that boots, configures, and manages every hardware block on the SoC. Your software runs as a shared library on embedded Linux, reaching into the chip to program PCIe links, initialize HBM controllers, configure PLLs, manage interrupt controllers, and orchestrate fabric interconnects across 270+ hardware block instances per chip — all deployed across millions of servers in AWS's global fleet.

Tech stack: C++17, CMake, GoogleTest, Python, SystemVerilog DPI, SPI, APB/AXI bus protocols, PCIe, UCIe, HBM, PLL, custom IPs

As the SoC Firmware Manager, you will:

- Manage, coach, and grow a team of 6 engineers — set technical direction, own hiring, and create an environment where strong engineers want to stay
- Coordinate deliverables across chip architects, RTL designers, verification engineers, validation engineers, and platform software teams — you're the single point of accountability for HAL readiness on every new chip program
- Own bring-up for new SoC tape-outs, from first-silicon power-on through production fleet deployment
- Prioritize work across multiple concurrent chip programs and customer teams, balancing urgent bring-up needs against long-term architecture investments
- Drive the architecture of our C++ template metaprogramming framework, BUTR (Built-in Unit Test for Registers), and HITL (Hardware-in-the-Loop) test infrastructure
- Ship the same C++ codebase to three execution environments: SystemVerilog DPI for chip verification, QEMU for emulation, and Carbon OS on embedded microcontrollers for production fleet
- Get into the weeds alongside your team — debug register-level HW/SW interactions, review code, and write code yourself when it matters

Most firmware teams target one platform and ship to a few thousand units. We target three platforms from a single source tree and deploy across AWS's global fleet — where a single register misconfiguration can impact millions of servers. Our software must be stateless, survive live-updates on running production servers without reboots, and be correct down to individual register bits. The microcontroller can reboot at any time — including during customer workloads — and the HAL must resume managing the SoC by querying hardware state on-demand. No cached state, no assumptions.

Your pre-silicon software runs in simulation and emulation months before real silicon arrives. When the chip comes back from the fab, you validate those predictions on real hardware — and when they don't match, you figure out whether it's a silicon bug or a software bug. For Trainium3, our HAL enabled a full ML training workload within 12 hours of first power-on: https://www.aboutamazon.com/news/aws/trainium-3-ultraserver-faster-ai-training-lower-cost

No ML background needed. Your firmware is the foundation that enables ML training across clusters of thousands of interconnected accelerators — you'll work on components like PCIe and HBM, but won't need to understand ML itself.

This role can be based in Cupertino, CA or Austin, TX. The team is split between the two sites.