About the role
You find out which hardware choices actually improve personal intelligence, by connecting real agent workloads to memory systems, accelerators and the software that uses them. The value of this role is in the measurement: most architectural intuition about AI workloads is a year out of date, and yours will be built from what our workloads really do.
The work
Profile representative inference, retrieval, training and agent workloads. Model compute, bandwidth, capacity and energy limits. Compare existing components before proposing specialized hardware, and partner with compiler and silicon teams on interfaces that can be evaluated early.
What good looks like
In your first 90 days, deliver a reproducible workload characterization and a quantified architecture recommendation with sensitivity analysis.
Evidence we look for
Bring computer architecture expertise and an ability to move between measurements, simulators and application behavior. Evidence may include research, silicon, systems prototypes or major performance work.
What we need to see
- Computer architecture expertise with evidence behind it: research, silicon, prototypes, or a major performance result
- You move fluently between measurements, simulators and real application behaviour
- You can tell a workload-driven conclusion from a benchmark-driven one
- You will say when the honest answer is that a hardware change is not worth it
Nice to have
- Memory-system or accelerator design specifically
- You have influenced a shipped silicon decision
- Published architecture research
The exercise
Explain when additional accelerator arithmetic throughput would fail to improve a memory-bound household workload.
Where and how we work
In the office together five days a week, in any of these cities. Remote-friendly around your family, arranged one person at a time.