ARNAV SINGH · DIGITAL DESIGN

INT8 compute,
verified from pins
to layout.

A signed multiply-accumulate core with 24-bit saturation, a working FIR application, and a measured protocol optimization. Built with open ASIC tools and explicit verification boundaries.

Baseline hardenedDSP gate-testedFused hardened experiment
Isometric rendering of the actual baseline MAC8 GDS

Actual baseline layout. Build 29401092054, SKY130, one Tiny Tapeout tile.

65,536
Signed operand pairs checked in RTL
200,000
Scheduled operations across two RTL variants
30.8%
Fewer simulated cycles per FIR output
15 / 15
Historical baseline prechecks passed
APPLICATION

Turn noisy samples into a useful signal.

128 outputs per variant. Zero integer mismatches.

Captured noisy input, MAC8 filtered output, and their frequency spectra

Host stores coefficients and sample history. MAC8 performs the arithmetic. The 16-tap filter introduces a 7.5-sample group delay. Plot values are from simulation; no physical throughput is claimed.

ARCHITECTURE

One multiplier. A precise contract.

Host
samples + coefficients
MAC8
8 × 8 → sat24
Readback
3 bytes

The host supplies signed operands over an 8-bit bus. A synchronized strobe accepts commands. Sticky overflow records saturation. The output register exposes the selected accumulator byte.

Driver and bring-up guide →
CYCLE EXPERIMENT

Three commands per fresh pair

364 cycles / FIR output

LDA → LDB → MAC. Includes CLR and full readback.

Same one-clock setup, three-clock strobe, three-clock low interval. 50 MHz assumed. Software GPIO overhead is excluded.

RESULTS

A tradeoff you can reproduce.

MetricBaselineFused candidate
Simulated cycles / 16-tap output364252
Equivalent outputs/s at assumed 50 MHz137,363198,413
Mapped standard cells, local synthesis805811
Mapped cell area, µm²6,152.26,231.0
Routed standard cells11881203
Routed standard-cell area, µm²10,547.60010,704.000
Worst setup slack, ns+1.556+0.702
Worst hold slack, ns+0.111+0.114
Worst-corner transition violations254282

Local mapping adds 1.28% cell area. Routed cell area adds 1.48% under identical flow configuration, PDK and source commit. Both meet setup and hold across nine corners, with zero DRC, LVS and antenna errors and 15/15 prechecks. Transition violations remain. The candidate needs a separate waiver review before adoption.

VERIFICATION

Prove the scope. Break the protections.

  • Exhaustive signed multiplication and exact FIR references.
  • Both saturation rails, sticky overflow, byte reads and reset recovery.
  • Nine application tests and 100,000 operations pass on the hardened candidate.
  • Fused scheduling invariants pass BMC and induction, with reachable covers.
  • Old-B, missing-MAC and stuck-low busy mutants are caught.
  • A missing pending-reset mutant fails the control proof.

Formal scope covers scheduling and selected handshake properties. Arithmetic equivalence and physical metastability are outside those proofs.

Requirements and proof boundaries →
EVIDENCE

Inspect the engineering.

Application results and source hashesCaptured input and output CSVSynthesis comparisonSame-flow routed comparison and corner metricsHardened candidate application captureHistorical raw-report manifest15 baseline prechecksOptimization method and resume wording

Baseline raw metrics: 1188 cells, +1.556 ns setup, +0.111 ns hold. Max-transition exceptions remain documented in W1. Portal submission status is unverified.