[lcus_masonry_article]<\/p>\n
In Project 18 \u2014 AutoPhi Future we talk about growing chips from a seed: one canonical blueprint from which every build, every foundry run, every tier is grown. Until now the seed was specs, RTL, and manifests. Now we have the first physical<\/strong> seed: a voxel built with OpenLANE2, real GDS and LEF, sitting in the repo and ready to be copied into every build.<\/p>\n That run is complete. The foundation is made.<\/p>\n The seed voxel is a single block \u2014 in this case We run the flow from WSL with a single script: When the script finishes successfully you have:<\/p>\n The build script already knows to include One definition, many harvests. The seed is the unit. We version the seed; we version the chips. From here we can grow: different foundries, different tiers, different nodes, but all from the same canonical voxel and the same flow. The first run is done. The foundation is laid.<\/p>\n Project 18 \u2014 AutoPhi Future. Seed voxel built with OpenLANE2; GDS and LEF in What the seed voxel is<\/h3>\n
autophi_voxel_blank<\/code> from our scale-ultimate RTL: a minimal footprint with the same interface as our hybrid voxels so it can slot into the grid. No logic inside; just the physical and electrical template. We run it through the full OpenLANE2 flow: synthesis, place and route, signoff, on SkyWater 130nm (volare). Out come GDSII and LEF. Those files go into 18-autophi-future\/seeds\/<\/code>. Our build script copies that folder into every build as Seeds\/<\/code> when it exists. So every package now has a real seed to grow from.<\/p>\nHow we got there<\/h3>\n
run_seed_openlane_wsl.sh<\/code>. It copies the voxel RTL and a minimal OpenLANE2 config into a design directory, runs OpenLANE2 (via Nix, Docker, or a local venv), then copies the final GDS and LEF into seeds\/<\/code>. We hit the usual bumps: the OpenLANE2 Nix flake runs Yosys tests in the build and they fail in the sandbox, so we apply a small patch to disable the check phase and the Nix build succeeds. The design has hundreds of IO pins, so we set the die large enough for them. The voxel is meant to be used inside a bigger design, so in a standalone run many pins are intentionally disconnected; we set ERROR_ON_DISCONNECTED_PINS: false<\/code> so the flow completes and we get the footprint. Once that was in place, the run went through to completion.<\/p>\nWhat you have when it’s done<\/h3>\n
\n
18-autophi-future\/seeds\/<\/code> \u2014 GDSII and LEF for the seed voxel.<\/li>\nscripts\/openlane_seed_design\/runs\/RUN_...\/final\/<\/code> \u2014 the same GDS\/LEF plus all intermediate artifacts.<\/li>\n<\/ul>\nseeds\/<\/code> in every build. So the package has a real physical seed: one voxel, one process node, one proof that the pipeline from RTL to GDS works.<\/p>\nWhy it matters<\/h3>\n
seeds\/<\/code>; the build adds it as Seeds\/<\/code>. One seed, many harvests.<\/em><\/p>\n