Machine Learning Research Engineer

Etched; Cupertino (CA); On-site; USD 120,000 - 160,000; Full time

Benefits offered by this job

Housing subsidy of $2,000/month

Daily lunch and dinner in the office

Relocation support for Cupertino

Job summary

A cutting-edge AI hardware company in Cupertino seeks a candidate to conduct research on custom AI chips, focusing on optimizing model performance and researching novel architectures. Responsibilities include collaborating on the software stack and interacting directly with hardware architects. Applicants should have a strong ML research background and experience with programming languages such as Python and Rust. This role emphasizes innovative solutions that are infeasible on traditional GPUs.

Qualifications

• Strong interest in HW co-design.
• Experience in distributed inference/training environments.
• Ability to conduct novel research.

Responsibilities

• Propose and conduct novel research on Sohu.
• Translate core mathematical operations into instruction sequences.
• Develop software performance for Sohu HW.

Skills

ML Research background

Experience with Python

Familiarity with transformer architectures

Tools

Pytorch

JAX

Rust

Job description

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. Etched Labs is the organization within Etched whose mission is to democratize generative AI, pushing the boundaries of what will be possible in a post-Sohu world.

Key responsibilities

• Propose and conduct novel research to achieve results on Sohu that are unviable on GPUs
• Translate core mathematical operations from the most popular Transformer-based models into maximally performant instruction sequences for Sohu
• Develop deep architectural knowledge informing best-in-the-world software performance on Sohu HW, collaborating with HW architects and designers.
• Co-design and finetune emerging model architectures for highest efficiency on Sohu
• Guide and contribute to the Sohu software stack, performance characterization tools, and runtime abstractions by implementing frontier models using Python and Rust.

Representative projects

• Propose and implement a novel test time compute algorithm that leverages Sohu’s unique capabilities to unlock a product could never be achieved on a typical GPU
• Implement diffusion models on Sohu to achieve GPU-impossible latencies that allow for real-time image generation
• Optimize model instructions and scheduling algorithms to optimize for utilization, latency, throughput, and/or a mix of these metrics.
• Implement model-specific inference-time acceleration techniques such as speculative decoding, tree search, KV cache sharing, priority scheduling, etc by interacting with the rest of the inference serving stack.

You may be a good fit if you have

• An ML Research background with interests in HW co-design
• Experience with Python, Pytorch, and / or JAX
• Familiarity with transformer model architectures and/or inference serving stacks (vLLM, SGLang, etc.) and/or experience working in distributed inference/training environments
• Experience working cross-functionally in diverse software and hardware organizations

Strong candidates may also have

• ML Systems Research and HW Co-design backgrounds
• Published inference-time compute research and/or efficient ML research
• Experience with Rust
• Familiarity with GPU kernels, the CUDA compilation stack and related tools, or other hardware accelerators
• 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 Cupertino

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 Cupertino, 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.

Equal employment opportunity

As set forth in Etched’s Equal Employment Opportunity policy, we do not discriminate on the basis of any protected group status under any applicable law.

Voluntary self-identification

For government reporting purposes, we ask candidates to respond to voluntary self-identification surveys. Completion of the form is voluntary, and any information provided will be kept confidential and used solely for compliance purposes. Participation or non-participation will not affect the hiring decision.