Extreme co-design
Also known as platform co-design, hardware-software co-design
In plain English
Co-design means developing connected parts of the system together so a gain in one part is not lost to a bottleneck elsewhere.
Technical definition
Extreme co-design is the Rubin article’s term for coordinated platform development across accelerator, host, interconnect, and networking products to serve the target workload.
Engineering details
A faster GPU can still wait for memory, communication, or request scheduling. Designing these components together changes the resources available to the serving stack. The article names six coordinated products: Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, and Spectrum-6.
Why it matters
The concept directs attention to whole-system performance rather than peak compute alone. It is an architectural approach, not a benchmark metric or an experimentally isolated explanation for a particular throughput multiplier.
How to read it in InferenceX
AgentX measures the resulting hardware-and-software configuration under multi-turn traffic. The Rubin article’s comparison does not independently vary all six products, so it cannot allocate the measured gain to each component or establish a universal co-design speedup.