CF-2: Eight Qwen Rows On Chip

Honest model_primitive scope: eight pinned tensor rows as on-chip ROM with row mux, not full-model silicon. OpenROAD placement through route, machine-checkable residency proofs, and portable handoff bundle for off-server ECC/signoff.

Key points

Eight real Qwen rows byte-verified and fixed in 128 generated weight-bank modules.

All eight rows pass RTL known-answer tests against software golden vectors.

Experimental GDS packaged for off-server ECC; foundry DRC/LVS not run on this server.

Rows

8 × 1024 INT4

in_proj_qkv rows including CF-1 anchor 6053.

Cells

16,035

Mapped ICsprout55 netlist after ECC dont_use policy.

DRT

0 violations

First-route pass on hardened DEF; no repair pass needed.

GDS SHA-256

47d14d05…

Pinned in designs/cf2/manifests/cf2_physical_artifacts.json.

Timing (exp.)

90 MHz

Placement-RC STA sweep; 100 MHz WNS −0.84 ns.

What CF-2 adds over CF-1

CF-2 proves the weight-bank and row-mux pattern scales from one row to eight without changing the honest claim boundary. Each row is independently verifiable; the top module selects among them with a 3-bit slot index.

Rows 6053, 5201, 650, 232, 3173, 6039, 6022, 2191 from in_proj_qkv.

36,994-instance hardened layout with tapcells, CTS, and fillers.

Mapped residency proof confirms all 128 banks reachable in the netlist.

Honest boundaries

CF-2 is experimental collateral, not tapeout-approved. Placement-RC STA closes at 90 MHz; 100 MHz WNS is −0.84 ns under the experimental RC model.

Run scripts/verify_cf2_gate.sh locally for full gate checks when tmp/ artifacts exist.

Transfer cf2_handoff_package.zip for ECC rtl2gds on a signoff host.

See docs/how_to/cf2_ecc_signoff_off_server.md for the off-server runbook.

Full Qwen model on chip remains out of scope.