Detailed routing is one of the slowest stages in an ASIC physical design flow. Even with TritonRoute’s multithreading support, routing large designs still becomes a bottleneck as complexity grows.
At WIOWIZ, we explored a practical question:
Can OpenROAD detailed routing be parallelized at the process level—without modifying TritonRoute or OpenROAD itself?
The answer is yes, with careful region partitioning and orchestration.
Why Multithreading Isn’t Enough
TritonRoute uses multithreading inside a single routing process. While helpful, it still means:
- One routing job
- One shared routing state
- Limited scalability on multi-core systems
True scalability requires multiple OpenROAD processes running in parallel, each routing a portion of the design.
Why Routing Is Hard to Parallelize
Routing is inherently global:
- Nets span across regions
- Power, ground, and clock networks are chip-wide
- Routing complexity is unevenly distributed
Naive region splits fail because wall-time is dominated by the most complex region, eliminating parallel gains.
Our Approach: OpenROAD as a Black Box
Instead of modifying TritonRoute, we built an external orchestration flow:
- Partition the design into routing regions
- Generate region-specific DEF and guide data
- Launch multiple OpenROAD processes in parallel
- Route each region independently
- Merge the routed outputs into a final DEF
Each OpenROAD process remains internally multithreaded, enabling process-level + thread-level parallelism.
The Key Insight: Balance by Complexity
Equal-area partitioning does not work. Routing difficulty does not correlate with geometry.
We achieved real speedups only after moving to complexity-aware region partitioning, ensuring that routing effort—not area—was balanced across regions.
Results
Using a RISC-V core in SkyWater 130nm:
- Sequential routing: ~644 seconds
- Parallel region-based routing: ~116 seconds
- ~5.6× speedup
All achieved without modifying OpenROAD or TritonRoute.
Why This Matters
This work demonstrates that:
- OpenROAD can scale beyond its default execution model
- Significant routing speedups are possible today
- External orchestration can extend open-source EDA tools safely
Full Technical Details
This dev.to post is a concise summary.
For full implementation details, partitioning strategy, iterations, and limitations, read the original article:
Parallel Region-Based Routing on OpenROAD
Top comments (0)