NoveetyAI
AI-native EDA for chiplets & advanced packaging

AI-powered multiphysics analysis, design exploration and sign-off

An agentic flow for chip integrity and reliability. Thermal, warpage, electromigration and power-integrity engines fast enough to run on every design candidate — each one checked against its own physics reference.

Explore the engines → Read the white papers →
ThermStack 3D temperature field of a 7 nm stepped HBM+CPU package WarpStack warpage map of a 2nm CoWoS package

Actual engine output: a ThermStack temperature field on a 7 nm stepped HBM+CPU package and a WarpStack warpage map of a 2nm CoWoS package.

Fast analysis

Full 3D thermal and warpage fields in milliseconds with AI, seconds with the numerical engine.

Design exploration

Sweep thousands of floorplans, stack-ups and materials in the time one 3D solve used to take.

Sign-off

Every fast answer is verified against a numerical ground-truth engine in the same tool.

The problem

Sign-off tools haven't kept pace with chiplet complexity.

In a chiplet package, electrical, thermal and mechanical physics no longer separate — and the tools accurate enough to trust are too slow to keep in the design loop. Why physics became a first-order constraint →

01 / Coupled physics

Physics is coupled

IR-drop drives heat, heat drives electromigration, CTE mismatch drives warpage. Checked one domain at a time, each answer misleads.

02 / Runtime

3D solvers are too slow

A full 3D solve costs minutes. A design-space search needs thousands of them, so physics gets checked last.

03 / Automation

Agents need callable physics

An autonomous design agent can only optimize what it can invoke. GUI-bound analysis can't close the loop.

How it works

An engineered loop, not a prompted one.

A human sets the objectives, the budget and the stopping condition once. After that the loop runs unattended — proposing candidates in parallel, evaluating each against real physics, and verifying with a separate agent before anything is signed off.

Engineered agentic loop: design, evaluate, verify, sign off A human sets objectives, budget and stopping conditions once. A loop controller driven by an LLM — Claude, Codex or an open-weight model — then runs unattended: it proposes design candidates in parallel, drives traditional EDA tools and the NoveetyAI physics engines to evaluate them, and hands the resulting margins to a separate verifying agent that checks them against the stopping condition on the numerical ground-truth engine. If the condition is not met, results and run history feed back through persistent memory into the next iteration; if it is met, the design proceeds to sign-off. OBJECTIVES · SET ONCE, BY A HUMAN targets · budget · stopping conditions MEMORY run history Pareto front what has been tried persists across runs LOOP CONTROLLER · RUNS UNATTENDED Claude / Codex / Open-Weight LLM proposes candidates · spawns parallel workers · reads margins no per-step prompting — the loop writes its own next move ① DESIGN — N CANDIDATES IN PARALLEL TRADITIONAL EDA Major commercial EDA flow OpenROAD synthesis · place & route · extraction · STA emits netlist, P/G grid, power maps NOVEETYAI ENGINES ThermStack — thermal WarpStack — warpage MetalStack — EM · TM · IR GridStack — transient IR-drop numerical + AI · CLI-first · structured JSON grid · power · geometry ② EVALUATE margins MARGINS worst-case droop · time-to-failure · hotspot temperature · peak-to-peak bow ③ VERIFY — MAKER ≠ CHECKER a separate agent tests the margins against the stopping condition — on the numerical ground-truth engine met ④ SIGN-OFF integrity + reliability OPTIMIZE — NOT MET, ITERATE

CLI-first

Every analysis is scriptable. Nothing in the loop needs a GUI.

Structured I/O

Machine-readable inputs; fields, margins and hotspots out.

Flow-agnostic

Plugs into any agentic orchestration layer you already run.

Closed-loop

Optimize against real physics, not a surrogate proxy.

The engines

Four engines enforce and gate the physics — AI accelerates them

Each engine pairs a physics-exact numerical solver with a fast path for exploration, on the same inputs and output format. Every analysis is a CLI call an agent can make thousands of times.

01 / Thermal analysis

ThermStack

Static and transient 3D temperature fields for 2.5D/3D chiplet stacks, read directly from IEEE 3Dblox.

182×vs. 3D FEM
10.6 msThermStack-AI per map
White paper → Details →
02 / Thermal-mechanical stress

WarpStack

Warpage maps and peak-to-peak bow for 2.5D/3D packages, from 3D sign-off to millisecond sweeps.

51 µsMOR-model per point
19.4 msWarpStack-AI per map
White paper → Details →
03 / Electromigration

MetalStack

Coupled electromigration, thermomigration and IR-drop over real power grids, with lifetime prediction.

EM+TM+IRcoupled solve
86×MetalStack-AI
Details →
04 / Power integrity

GridStack

Transient power-grid simulation and dynamic IR-drop on million-node designs. Integrated with Ergodex AI as a callable engine for its agent-driven design experiments.

11.9×vs. direct solve
<6e-6max normalized error
Details → Ergodex AI integration →
Results

Measured against full 3D references.

Every fast path is scored against a numerical ground-truth engine on the same geometry, so each answer comes with a known error.

182×faster than 3D FEMThermStack, 0.509 K mean RMSE across eight native IEEE 3Dblox packages.
0.036 µmfrom the 3D warpage solveWarpStack MOR-model on a 2nm CoWoS package, at 798,000× the speed.
0.32 KAI error over 360 unseen casesThermStack-AI full 3D maps in 10.6 ms, validated against FEM-3D.
White papers

The technical evidence, one paper at a time.

This page stays short on purpose. Methods, benchmarks, figures and their scope notes live in a growing series of white papers you can read, share and cite in a design review.

Thermal analysis

ThermStack: geometry-resolved static and transient thermal analysis

0.509 K from 3D FEM at up to 182× the speed across eight IEEE 3Dblox packages, plus a graph-neural surrogate at 10.6 ms per map.

Read the paper →
Thermal-mechanical stress

WarpStack: warpage sign-off and design-space exploration

3D, reduced-order and graph-neural methods from one design description; within 0.036 µm of 3D at 798,000× the speed on a 2nm CoWoS package.

Read the paper →
Start with your package

Put integrity back in the loop.

The NoveetyAI suite is in active development. Request access or a walkthrough on your own 2.5D/3D chiplet designs.

Request access Book a demo →