Inference Software Engineer - Collectives

About Etched

Etched is building the world's first AI inference system purpose-built for transformers - delivering over 10x higher performance and dramatically lower cost and 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. Backed by hundreds of millions from top-tier investors and staffed by leading engineers, Etched is redefining the infrastructure layer for the fastest growing industry in history.

Job Summary

Etched's Inference SW team enables optimal mapping of models to Sohu's dataflow architecture and serving requests across multiple chips, hosts and racks. We are seeking a highly skilled and motivated engineer to formalize and optimize our collectives (e.g. Send / Recieve, AllReduce, Broadcast, etc.). You'll build SW enabling frontier inference performance to satisfy exponentially growing serving demand.

In this role, your core focus will be working across systems and research to realize Mixture of Expert (MoE) architectures on Sohu's system. You will play a key role in scaling out Sohu's nascent runtime, with a focus on collectives.

Key responsibilities

• Formalize and optimize our collectives (e.g. Send / Recieve, AllReduce, Broadcast, etc.)
• Collaborate across systems and research teams to bring MoE architectures to Sohu's runtime
• Optimize expert routing and communication layers using Sohu's collectives
• Contribute to scaling and enhancing Sohu's runtime, including multi-node inference, intra-node execution, state management, and robust error handling
• Develop tools for performance profiling and debugging, identifying bottlenecks and correctness issues

You may be a good fit if you have

Strong candidates may also have experience with

Benefits

Compensation Range

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.