-
Notifications
You must be signed in to change notification settings - Fork 0
Tapeout Checklist
Tapeout is the moment the design is released to the foundry. After it, nothing can be changed — a mask set costs real money and a respin costs months. This page is the gate: every item must be checked, evidenced, and signed by a named owner before release.
Treat it as a checklist in the aviation sense. Not a summary of what you probably did, but a list you physically walk through with someone else watching.
- Freeze before you check. Any change after a check invalidates it.
- Evidence, not memory. Every item needs an archived log or report.
- Named owner per item. "The team checked it" means nobody checked it.
- Two people per critical item. The person who did the work and one who didn't.
- Waivers are documented or they don't exist. Justification, reviewer, date.
- If you don't understand it, it isn't signed off.
flowchart LR
A[Design freeze] --> B[Full signoff run]
B --> C[Checklist review]
C --> D[Fix + re-verify]
D --> C
C --> E[GDSII release]
E --> F[Foundry acceptance]
style C fill:#2d6a9f,color:#fff
The C → D → C loop is where most tapeouts spend their final weeks. Budget for it — first-pass checklist review always finds something.
| Check | Criterion | Evidence |
|---|---|---|
| DRC clean | Zero violations or documented waivers | Pegasus report |
| LVS clean | No discrepancies | Pegasus report |
| ERC clean | No floating gates, no supply shorts | Pegasus report |
| Antenna | Zero violations | Pegasus report |
| Metal density | Within min/max per layer, fill inserted | Density report |
| Seal ring | Present and correct per foundry spec | Layout inspection |
| Scribe / dicing | Correct per shuttle spec | Layout inspection |
See Signoff for how these are run.
| Check | Criterion |
|---|---|
| Setup | WNS ≥ 0, all views, extracted parasitics, SI on |
| Hold | WNS ≥ 0, all views — no exceptions |
| Design rules | Zero max_tran, max_cap, max_fanout violations |
| Analysis coverage | ~100%, no silently unconstrained paths |
| Exceptions reviewed | Every false path and multicycle path justified |
| All modes | Functional, scan, at-speed all pass |
| Clock domain crossings | Declared and verified |
Hold is non-negotiable — see Timing Closure. A hold violation is a dead chip at every frequency.
| Check | Criterion |
|---|---|
| Static IR drop | ≤ 2–3% VDD |
| Dynamic IR drop | ≤ 5% VDD, with a real activity trace |
| Electromigration | Within foundry limits for target lifetime |
| PDN connectivity | Every macro and cell connected; verify_power_via clean |
| Decap | Inserted, ratio within plan |
| Power domains | Match UPF; isolation and level shifters correct |
See Power Planning.
| Check | Criterion |
|---|---|
| LEC | Final netlist equivalent to RTL, all points mapped |
| Gate-level simulation | Passes with SDF back-annotation |
| Regression | Full Verification suite green on the tapeout RTL |
| Reset | Verified from all reset states |
| Scan chains | Continuity verified, chain order documented |
| ATPG coverage | Meets target |
| Check | Criterion |
|---|---|
| Padring | Complete, correct pad order, no gaps |
| Pad assignment | Matches package/bond diagram exactly |
| ESD | Protection on every pad; correct clamps |
| Power pads | Sufficient count and distribution for peak current |
| I/O configuration | Drive strength, pull-ups, levels correct |
| Corner cells | Present |
| Bond pad opening | Matches packaging requirement |
Pad assignment errors are among the most common and most fatal student-tapeout mistakes. Check it against the bond diagram, pad by pad, with two people.
| Item | Notes |
|---|---|
| GDSII | The layout itself; correct top cell name |
| LEF | Abstract, if the block is reused |
| Netlist | Final routed Verilog |
| SDF | Timing for gate-level simulation |
| SPEF | Extracted parasitics |
| Timing reports | All views, archived |
| DRC/LVS reports | Clean, archived |
| Waiver document | Every waiver with justification and reviewer |
| Bond/package diagram | Pad-to-pin mapping |
| README | How to reproduce the whole flow |
Archive enough that someone can rebuild this exact database in a year:
- Tool versions (exact, including patch level)
- PDK version and rule deck versions
- All scripts, constraints, and configuration
- Git commit hash of the RTL
- Full logs from every flow stage
- The final database itself
Our chamber pins tool versions deliberately for exactly this reason — record which ones you used.
| Item | Owner |
|---|---|
| Design frozen, no pending ECOs | PD lead |
| All checklist items evidenced | PD lead |
| Waivers reviewed and approved | PD lead + advisor |
| Foundry/shuttle submission requirements met | Integration owner |
| Submission deadline confirmed | Program owner |
| Backup of the complete database | Anyone — but verify the backup restores |
| Mistake | Consequence |
|---|---|
| Late ECO after signoff | Invalidates completed checks; the classic respin cause |
| Pad assignment mismatch | Chip cannot be bonded correctly — unrecoverable |
| Hold violation waived under schedule pressure | Dead silicon at every frequency |
| Signed off on static IR only | Dynamic hotspot fails in the lab |
| One precision/functional mode unverified | Half the chip untested |
| Undocumented waiver | Nobody can tell later whether it was safe |
| Wrong GDS top cell name | Foundry rejects the submission |
| Missing fill | Density DRC failure at the foundry, after your deadline |
| No archive | Cannot reproduce or debug the silicon you receive |
Lambda's chipathon target hardens the decode attention datapath — Q·Kᵀ → softmax → P·V, plus KVE, TIU, and the ACU gate. Specific items to add to the generic list:
- Verify both precision modes. The INT8/FP16 precision controller means two distinct datapath configurations. Timing, power, and functional verification must cover both. Signing off one mode is the highest-risk shortcut available here.
-
KVE SRAM connectivity, pin by pin. Macro power/ground pins are a recurring LVS failure. Check every bank explicitly rather than trusting
route_special. - Dynamic IR with a real decode trace. The MatE array's sustained uniform switching is the worst case, and only a realistic workload trace exposes it. Get one from Verification, not a synthetic pattern.
- Clock domain crossings between KVE and the compute array — declared, verified, and covered by STA.
- Array boundary DRC. The operand-broadcast and result-drain regions at MatE's edges are the congested areas from Routing and therefore the likeliest DRC hotspots.
- Metal density across the array. A uniformly dense regular array plus sparse surroundings hits density limits at both ends.
- Reference model parity. Confirm the taped-out RTL still matches the Python golden model. The repo's lab-notebook standard requires this evidence recorded, not asserted.
- The shuttle deadline is immovable. Unlike an internal schedule, a shuttle slot does not slip for you. Work backwards from the submission date and put the signoff freeze at least two weeks earlier than feels necessary.
-
GF180MCU submission — the shuttle organiser will run their own DRC/LVS on your GDSII and reject it if unclean. Their acceptance is the real gate, not your local run. Full-chip assembly is in
chip/pdk/gf180/. - Padring first, not last. Pad count and pitch can force the die to grow independently of core area. Discovering this late is a common and expensive student mistake — check it during Floorplanning.
- Do a dry-run tapeout. Take a trivial design — an inverter ring oscillator — all the way through DRC, LVS, and GDS submission format early in the project. It surfaces flow and submission problems while there is still time to fix them, and it is far cheaper than discovering them with the real design.
- Two people, out loud. Walk this checklist with a partner reading items aloud. It catches things silent review does not.
- Foundry / shuttle submission guidelines — the authoritative requirements list; read it early, not at the deadline
- Cadence Pegasus, Tempus, Voltus user guides — how each signoff check is run
- Efabless / SSCS Chipathon documentation — open-shuttle submission requirements and precedents
- SkyWater and GF180MCU PDK documentation — rule decks and required deliverables
- Kahng, Lienig, Markov, Hu — VLSI Physical Design — for the concepts behind each check
See also: Signoff · Timing Closure · Power Planning · Floorplanning · Physical Design