Khan, M. (CSE) – Clock Mesh Analysis and Optimization using Surrogate Models with Graph Neural Networks

Clock Distribution Networks (CDNs) are an essential part of every modern chip that
distributes a clock to all the synchronous elements. Among various CDN architectures,
clock meshes are among the most robust architectures to On-Chip Variation (OCV) and
provide low skew due to its redundant paths, thereby allowing high frequency opera-
tion. However, the trade-offs are: more routing resources, increased power consumption,
and the inability of Static Timing Analysis (STA) to annotate delays due to no single
source-to-sink path. Therefore, a full SPICE simulation is the only option for timing
analysis of clock meshes. Existing clock-mesh synthesis frameworks incorporate heuris-
tics to avoid SPICE simulations, however, they are inaccurate especially under OCV.
In this thesis, a solution for faster and accurate analysis of clock meshes under OCV
is proposed. This is the first work on representing a clock mesh in graph domain and
training a GNN-based surrogate model to predict timing and power metrics of the clock
mesh. Furthermore, this surrogate model is used inside an optimization loop to buffer
a given clock mesh considering OCV which has never been feasible due to SPICE being
expensive and heuristics being inaccurate. The proposed surrogate model achieves aver-
age errors of 3.40 ps in delay, 4.14 ps in slew and 4.46 ps in skew as compared to SPICE.
Whereas, the heuristic method has 74.37 ps delay error and 27.51 ps skew error, while
it cannot annotate slews. The surrogate model predicts power within 3.4% of SPICE
on average, while delivering 18×-1706× speed-ups over the approximate method under
OCV modeling, enabling faster and more accurate clock-mesh design-space exploration.
When used inside the optimization loop the surrogate model finds buffering candidates
that, at a balanced tradeoff, improves 99th-percentile skew by 49.1%, reduces power
by 25.7%, and reduces area by 37.3% at the cost of a 6.8% increase in 99th-percentile
maximum slew as compared to the state-of-the-art work.
Event Host: Muhammad Hadir Khan, PhD Candidate, Computer Science & Engineering
Advisor: Matthew Guthaus