ML Performance Characterization Engineer

About Etched

Etched is building AI chips that are hard-coded for individual model architectures. Our first product (Sohu) only supports transformers, but has an order of magnitude more throughput and lower latency than a B200. With Etched ASICs, you can build products that would be impossible with GPUs, like real-time video generation models and extremely deep & parallel chain-of-thought reasoning agents.

Job Summary

Join our team as a Senior ML Performance Characterization Engineer and take the lead in illuminating the performance landscape of our cutting-edge ML accelerator. We are seeking a highly skilled engineer to design and develop a sophisticated performance analysis tool, tailored specifically for Sohu. You will be instrumental in creating the essential tooling that enables our ML engineers and customers to understand workload behavior, identify performance bottlenecks, and unlock the full potential of Sohu accelerating the most demanding ML applications in the world. This is a unique opportunity to shape performance analysis for novel hardware from the ground up.

Key responsibilities

• Tool Architecture & Design: Lead the design and architecture of a comprehensive performance analysis suite, including data collection mechanisms, data processing pipelines, analysis engines, and user interfaces (CLI and/or GUI).

• Low-Level Data Collection: Develop robust methods to capture performance data directly from our custom ML accelerator hardware (e.g., hardware performance counters, execution unit status, memory access patterns) via driver interfaces or other mechanisms.

• Host & System Tracing: Implement tracing for host-side API calls (runtime libraries, driver interactions) and system-level events (CPU activity, PCIe traffic, memory usage, network contention) related to Sohu workloads.

• Data Correlation & Synchronization: Design and implement techniques to accurately correlate performance events across the host CPU, device driver, PCIe bus, and multiple accelerators, ensuring precise time synchronization.

• Performance Analysis Engine: Build analysis modules to automatically interpret collected trace and counter data, identifying key performance limiters (e.g., compute-bound, memory bandwidth-bound, latency-bound, PCIe-bound, specific hardware bottlenecks).

• Visualization & Reporting: Develop intuitive visualizations (timelines, dependency graphs, resource utilization charts, statistical summaries) to clearly communicate performance characteristics and bottlenecks to users.

• Collaboration & Support: Work closely with hardware architects, firmware engineers, driver developers, compiler engineers, and ML application engineers to understand their needs, define tool requirements, and provide expert guidance on performance analysis and optimization using the tool.

Representative projects

• Architect and implement the core data collection framework for hardware performance counters on a custom PCIe-based accelerator.

• Develop a kernel driver module or user-space service for low-overhead tracing of accelerator activity.

• Design and build a correlated timeline view visualizing CPU API calls, driver submissions, PCIe transfers, and accelerator execution units.

• Create an analysis pass to detect and quantify memory access inefficiencies or PCIe bandwidth saturation while transacting on a PCIe-attached accelerator.

You may be a good fit if you have

• Strong proficiency in C/C++ and Python.

• Deep understanding of computer architecture (CPU, GPU, accelerators), memory hierarchies (caches, DRAM), and interconnects (especially PCIe).

• Proven experience in low-level performance analysis, profiling, and bottleneck identification on complex hardware systems (GPUs, CPUs, FPGAs, or custom ASICs).

• Experience with performance analysis tools (e.g., NVIDIA Nsight, AMD uProf, Intel VTune, perf, Tracy, ETW).

• Solid understanding of operating system internals (Linux preferred), including scheduling, memory management, and driver interaction.

• Experience working close to hardware, potentially reading performance counters or interacting directly with device drivers.

Strong candidates may also have experience with

• Direct experience developing performance analysis or debugging tools.

• Experience with ML accelerator architectures (GPUs, TPUs, etc.).

• Experience with kernel-mode driver development (Linux or Windows).

• Understanding of compiler internals, code generation, and optimization.

• In-depth knowledge of the PCIe protocol and analysis tools (PCIe analyzers).

• Experience with firmware or embedded systems development.

• Experience with hardware description languages (Verilog, VHDL) or hardware verification.

Benefits

• Full medical, dental, and vision packages, with 100% of premium covered

• Housing subsidy of $2,000/month for those living within walking distance of the office

• Daily lunch and dinner in our office

• Relocation support for those moving to West San Jose

How we’re different

Etched believes in the Bitter Lesson. We think most of the progress in the AI field has come from using more FLOPs to train and run models, and the best way to get more FLOPs is to build model-specific hardware. Larger and larger training runs encourage companies to consolidate around fewer model architectures, which creates a market for single-model ASICs.

We are a fully in-person team in West San Jose, and greatly value engineering skills. We do not have boundaries between engineering and research, and we expect all of our technical staff to contribute to both as needed.