The posting, in SambaNova's own words
archived Oct 22, 2026SambaNova is a leader in next-generation AI infrastructure, delivering a full-stack inference platform for customers worldwide. At the core of SambaNova's technology is the RDU (Reconfigurable Dataflow Unit) — a chip built on a dataflow architecture rather than the traditional GPU model. Its decode performance is especially strong for agentic workloads like multi-turn agents, code generation, and long-running applications. RDUs are packaged into SambaRack, rack-scale hardware that lets customers deploy state-of-the-art models with better performance, greater energy efficiency, and faster time to value.
About the Team
The Inference Systems Performance team answers how fast SambaNova's systems can serve large language models and what it takes to get there. We capture real production traffic, benchmark it faithfully, model and simulate configurations that do not exist yet, and profile the distributed serving pipeline across host, accelerator, and fabric. Our work feeds serving optimization today and hardware and capacity planning for the next generation. We sit close to model optimization, systems, hardware, and product, and the whole company depends on our numbers.
Read the full posting ↓
About the Role
You will be the architect for end-to-end inference performance: how a request moves through tokenization, prefill, decode, and the fabric between them, and how a deployment is sized against customer SLOs. The work has two coupled pillars. One is reproducible workload capture and benchmarking, building replayable representations of real and increasingly agentic traffic so measurements reflect production rather than a naive load script. The other is performance modeling and simulation, turning measurement into a what-if capability for configurations and hardware that do not exist yet. The frontier you will help define is heterogeneous, disaggregated inference, with GPU on prefill and the RDU on decode. That opens hard problems in networking, storage, prompt caching, and tail-latency-bound data movement. Inference systems performance is a young field and most answers are still being discovered, so you will spend your time on ambiguous problems with no established solution, and you will set the technical direction that others build on.