F

FuriosaAI

Power-efficient AI inference accelerator chips

Scale-Up Ethernet Engineer

Hardware Systems IntegrationPosted on Nov 12
Full-timeHsinchu, North District, Hsinchu City, TaiwanNo salary listedIn-person

About the Job

Requirements

  • A BS in Computer Science, Electrical Engineering, or a related field, or equivalent practical experience.
  • Understand data-center Ethernet networks, including switches and lossless/RDMA networking, and have experience configuring and debugging network devices or systems on Linux.
  • Have experience developing or debugging semiconductor system software that works with hardware at register and memory-map level, such as SoC firmware, device drivers, silicon bring-up, or diagnostic software.
  • Have hands-on experience with bring-up, integration testing, or performance analysis of rack- or cluster-scale systems combining servers, switches, and accelerators.
  • Be able to automate test execution, device control, and log collection with Python or shell scripts.
  • Actively use AI tools, including LLMs and coding agents, in day-to-day work and verify their output against evidence.
  • Be able to form hypotheses from packets, logs, and performance counters, and narrow down root causes through reproducible experiments.
  • Be able to collaborate with engineers from different disciplines and document test methods and results clearly.

Responsibilities

  • Validate the scale-up Ethernet endpoint controller IP integrated into the accelerator SoC from pre-silicon prototypes through silicon. Understand its interaction with the SoC on-chip interconnect (AXI), DMA, memory map, interrupts, and firmware, and follow problems beyond the link into the chip, resolving them with RTL and firmware engineers.
  • Configure NICs and switches based on traffic profiles, understand switch NOS behavior, validate link-layer functions including buffers, queues, PFC, LLR, and CBFC, and evaluate purpose-built scale-up switches to improve performance and stability under congestion.
  • Reproduce data-transfer errors and performance bottlenecks using packet captures, device counters, logs, and telemetry, and isolate causes to hardware, firmware, drivers, software, or fabric.
  • Build clusters running collective communication such as all-reduce and all-to-all, plus LLM inference and training traffic, to find issues and corner cases missed by lab tests. Measure bandwidth, latency including tail latency, and stability; verify configuration changes and detect performance regressions.
  • Analyze MAC, PCS, FEC, and SerDes status and error counters; validate link stability and interoperability across optical modules, DAC/AEC/AOC cables, and switch ports. Use CMIS to check module application mode, firmware version, temperature, and optical power, and verify stability through repeated resets and temperature changes.
  • Bring node BMCs, switches, power shelves/BBUs, PDUs, and cooling under one rack manager that handles inventory, power sequencing and remote control, and firmware updates. Use Redfish and PLDM for nodes and gNMI/YANG for switches, define the API for the upper-level orchestrator with reference to OCP specifications such as OpenRMC, and design scale-up fabric management for endpoint identity and address assignment, topology discovery, link bring-up ordering, and domain partitioning.
  • Observe power, thermal, per-lane FEC link, optical-module, and device health as time series in one place; design failure detection, isolation, fail-safe recovery scenarios, and RAS; and verify them with fault injection.
  • Build racks from compute trays with NPUs and NICs, switch trays, power shelves, management networks, and connecting DAC/AEC/AOC and optical cabling. Validate progressively from link bring-up through full rack integration, and lead the proof of concept to move the current rack validation environment to OCP-standard configurations such as Open Rack.
  • Automate remote console and PDU control, test execution, and result collection; build dashboards and technical documentation for long-duration stability testing and regression validation.
  • Delegate device bring-up, builds, test execution, log analysis, and documentation to AI agents such as Claude Code and Codex. Design equipment reservation and locking, remote-control APIs, and job queues so multiple agents can share lab equipment without conflicts, and check agent output against evidence.
  • Bring developments from OCP (ESUN, Open Rack, Hardware Management), UEC, the SONiC Scale-Up Working Group, and other standards bodies into designs, and work with switch, power, BMC, and test-equipment vendors.

Desired Qualifications

  • Experience with link bring-up, performance evaluation, or interoperability testing in 100G–800G high-speed Ethernet environments.
  • Understand or have hands-on experience with MAC/PCS/FEC/SerDes, QSFP-DD/OSFP optical modules and DAC/AEC/AOC cables, and CMIS-based module management.
  • Experience designing rack-scale copper (DAC/AEC) and optical (AOC and optical-module) cabling, or qualifying cables and modules, including interoperability testing across vendors and manufacturing lots.
  • Experience configuring lossless Ethernet fabrics and tuning congestion control, including PFC, ECN, buffers, and queues.
  • Understand switch NOS architecture and behavior, and have experience with Broadcom Tomahawk-family switches.
  • Experience analyzing performance bottlenecks in DMA, remote memory access, and collective communication such as all-reduce and all-to-all on NPU/GPU accelerator systems.
  • Experience validating interaction between Ethernet controller IP and the SoC, including on-chip interconnect, DMA, and memory subsystem, or tracking down issues at the C/C++ and RTL level with driver, firmware, and hardware engineers.
  • Experience with rack and server management using BMC/OpenBMC, Redfish, PLDM/MCTP, or OpenRMC for inventory, firmware updates, power and thermal control, and failure recovery, or experience developing and operating a rack manager or fleet management system.
  • Experience building hardware test labs, CI-based regression testing, telemetry, or long-duration stability-test infrastructure.
  • Experience building and running workflows that split work across multiple AI agents, such as device-control APIs for agents, session and job management, and result-verification harnesses.
  • Experience with OCP rack hardware such as Open Rack (ORv3/ORW), 48 V and 800 VDC power, power shelves, and BBUs, or bringing up racks with ODMs.
  • Understand scale-up standards such as OCP ESUN/SUE-T, UEC (LLR, CBFC), and UALink, or participate in OCP or SONiC communities.

About the Company

FuriosaAI

Power-efficient AI inference accelerator chips

FuriosaAI develops AI accelerators for data centers. Its flagship RNGD chip is a power-efficient accelerator designed to boost performance per watt for running large language models and other demanding AI workloads. The hardware is configurable with multiple processing elements to fit different workloads, and it is paired with advanced compiler technology and profiling tools to identify bottlenecks and optimize efficiency. This combination of flexible hardware and software optimization differentiates FuriosaAI from competitors, focusing on maximizing performance while keeping energy use low. The company targets data centers and enterprises that need high-performance, energy-efficient AI solutions, and aims to make AI computing more sustainable and scalable by selling its chips and related software tools and pursuing long-term contracts with large-scale customers.

Company size
201-500
Company stage
Series D
Total funding
$887.5M
Headquarters
Seoul, South Korea
Founded
2017

Recently Posted Jobs

Explore more opportunities at FuriosaAI

See more jobs