1. Purpose of the gate
The models and workloads the design is judged against are chosen, characterised and agreed by product, architecture and software leads.
This gate fixes the problem the accelerator solves. After it, performance modelling, architecture and the compiler are all built around the agreed set, so an error here propagates through the whole lifecycle. The review checks that the set is representative of the target market, characterised well enough to model, and small enough to analyse in depth.
It also checks that the assumptions about where models are going are written down with an owner, because the set will be revisited at later gates as models change.
A good workload set is a deliberate portfolio. It mixes the workloads that bring revenue at launch with at least one that represents where models are heading, and it describes each with enough precision that two engineers modelling it independently would reach similar conclusions. The review tests both qualities.
The set also becomes a shared language. Architects, compiler engineers, benchmarking engineers and applications engineers will refer to the same workloads by the same names for years, so names, versions and input conditions are fixed in the baseline.
10 more sections for registered members.