Skip to content

Latest commit

 

History

History
50 lines (40 loc) · 2.5 KB

File metadata and controls

50 lines (40 loc) · 2.5 KB

Full-Chip Verification Plan

Executable stack

make phase9 runs standard-library unit/model tests plus the assertion-enabled full-chip Verilator suite. The suite combines the long directed smoke test with 1,024 deterministic randomized commands, randomized completion backpressure, multiple reset lengths, coverage collection and a 90% executable line-coverage gate. Block regressions remain part of local-regress.

The trace sim/workloads/phase9_full_chip_trace.json represents one reused-KV agentic token flow across KV mapping, DMA, dependency scheduling and a local collective. dv/phase9/invariants.py independently checks command/completion conservation, tenant KV isolation, DMA apertures, refcounts, dependency issue, partition routing and the complete 1,280-bin required cross.

Critical assertions

Contract Executable checker
No cross-tenant KV access FullChipMonitor.read_kv negative test
No DMA outside range RTL directed test plus monitor negative test
Exactly one completion per accept rcif_top_assertions and scoreboard
No refcount underflow monitor negative test
Dependencies ready before issue scheduler RTL/test and monitor
No collective partition escape monitor negative test
No deadlock under legal backpressure exhaustive model and random RTL stalls

Coverage and signoff

  • Functional: all required cross bins must be hit; no missing bin is waived.
  • Code: bounded open-source RTL line coverage is at least 90%. Production interfaces and macros are added to the denominator when they exist; exclusions are recorded in coverage_waivers.md.
  • Assertions: every critical contract has a positive or negative executable test. RTL-port omissions are tracked, not silently treated as proven.
  • Seeds: deterministic default seed is mandatory in CI; additional seeds may be layered without changing the reference result.

Reset, CDC/RDC and gate flow

Current RTL has one clock and one asynchronous active-low reset. Lint plus random reset sequences close the bounded single-domain scope. Actual gate-level SDF simulation, scan reset, memory macro timing, CDC/RDC reports and X-prop are hard signoff gates once a standard-cell/memory library and synthesized netlist exist. They are not claimed by the RTL surrogate.

The bounded Phase 9 implementation is complete when make phase9 and make local-regress pass with no assertion or coverage failure. Tapeout-grade signoff remains conditional on closing the explicit production-library gates.