Keynote II
summary
Claim: on-chip wires dominate energy and latency, so a 2D mesh (add diagonals if possible) plus separate request/reply networks wins. Router design uses simple direction-order routing and two‑phase request/ack flow control with no virtual channels, yielding about 100 ps input‑to‑output router delay and wire‑dominated nanosecond tile transit. Circuits reuse supply to cut wire energy ~4x in silicon prototypes taped into a GPU shuttle, and embedded AND‑tree barriers give 20–30 ns global syncs on large chips.
AI-written summary · about
chapters (14)
- 0:00Start
- 10:10Network Design
- 18:28Asynchronous Flow Control
- 26:38Summarize the Router Design
- 31:13The Elastic Buffer Flow Control
- 38:27Crosstalk
- 41:18Fast Barrier Network
- 44:26Half Cycle Synchronizers
- 45:47Future Directions
- 48:50Optimum Repeater Delay
- 49:23What Is the Maximum Achievable Signal Velocity for these 80 Nanometer Pitch Wires
- 51:42Standardized Interfaces
- 52:22Barriers to Entry
- 55:47Low Energy Circuits To Drive the Wires
