Video Summary - Building the AI-native electronic design flow

Summary of technological concepts, product features, and analysis (Siemens x Nvidia DAC livestream)

Framing: “AI-native chip/PCB design” as “agentic engineering”

The discussion frames the goal of EDA engineers as solving real problems (rather than writing RTL/Verilog directly). It then introduces AI agents as the mechanism to carry out multi-step EDA workflows.

Key points:

Agents as partners to engineers: they run tools, perform verification loops, and accelerate design cycles.
Problem execution, not LLM guesswork: the agent works through tool-driven steps that validate outputs.

Three pillars of the Siemens–Nvidia AI-native EDA approach (CES announcement)

Faster engines
- Driven by GPU acceleration and AI techniques embedded inside EDA tools.
- Target outcome: faster turnaround across semiconductor and PCB design steps.
Smarter execution + trusted outcomes
- Agentic AI can autonomously run tasks across a workflow, including:
  - digital verification
  - high-level synthesis
  - customized C design
  - design-for-test
  - place & route
  - physical verification
- Agents are emphasized as self-verifying, using tight tool loops rather than relying on LLM predictions.
Trusted productivity at high cost-of-failure
- Because semiconductor failures can be extremely expensive (tens of millions), verification and reliability are central to adoption.

Long-running agents (milestone discussed, tied to “2026”)

A “long-running agent” is defined as one that:

maintains state/context across multiple tool calls
performs reasoning between steps to decide what to do next
calls different tools while preserving progress (not simply “running for a long time”)

Enablers:

very fast accelerated tools (to support rapid iteration)
agent orchestration + stateful context

Nvidia infrastructure stack for agentic EDA

The stream highlights a multi-layer Nvidia stack:

CUDA-X libraries and new libraries (mentioned: CUIS(S)) to accelerate core chip/system workloads.
NEMO / Physics Nemo libraries for building and evaluating agentic “AI physics” models and reasoning components.
Open Shell referenced for secure routing/selection of model sizes based on prompt needs.
Nemotron referenced as a secure runtime/reasoning model approach.

Important clarification:

The agent/LMM is not running “ground-truth physics” directly.
Ground truth comes from the EDA engines/tools that the agent invokes.

Siemens EDA: “Fuse EDA AI Agent” vs simply “layering AI”

Siemens positions the effort as building the full stack, not just adding AI on top of existing flows:

GPU-accelerated EDA engines
Long-running agent orchestration across the Siemens EDA portfolio

Reported accelerator impacts (quoted examples):

Solido Spice simulator: 10x speed
parasitic extraction: 10x+
Calibre OPC parasitic/related jobs: 50x+

Long-running agent integration uses Nvidia-related components:

NEMATRON/NEMO-related tooling for agent optimization/evaluation
Nemo Gym
Open Shell / secure model selection

Productivity and token-cost claims

A showcased workflow:

library characterization described as a “several-week job”

Claims mentioned:

agents provide about 10x productivity improvement
token efficiency / cost reduction (~10x token cost reduction)

Emphasis:

avoid large “token bills” even when agents run autonomously.

Expected engineering impact (time, output, design quality)

Design cycle time reduction targets

Customers aim for about 50% shorter design cycle times, with examples such as:

9 → 5 months
12 → 6 months

Quality improvements

Agents enable:

deeper design exploration
pushing limits across power/performance/area/yield compared with manual tool interaction

Workforce leverage

Agents help junior and mid-level engineers operate faster by reducing the need to be tool experts. This is framed as important given:

anticipated engineer shortages
rising industry complexity

Cost and scalability discussion: efficiency beyond raw model usage

Both Siemens and Nvidia address cost via a value-first framing:

if outcomes improve (time-to-market, quality), token cost becomes less dominant

They still emphasize:

token efficiency and model right-sizing
- model routing/orchestration: use larger models for complex prompts; smaller models for micro-tasks
- Siemens mentions skills files and “agent infrastructure” to optimize prompts and model selection
Nvidia emphasizes that orchestration is under-discussed and crucial for agent reliability and efficiency at scale

Trust, verification, and “self-verifying” agents

A major trust claim:

Ground truth comes from EDA tools, not the LLM’s predictions

Operational model:

the agent runs a loop:
- tool execution
- data generation
- tool-based verification
- adjust/retry if off-track

Additional Siemens requirement:

expose richer APIs from EDA engines to agents (agent interface differs from human UI)
agents should not be a black box; engineers must see:
- what steps the agent took
- where it converged or deviated

Institutional knowledge:

engineers can intervene and store knowledge as skills
persistence/improvement described as capturable in a central data lake (Fuse Agent infrastructure)

Security and sandboxing challenges (deep dive)

Unlike chatbots (typically web endpoints), EDA agents face:

sensitive, IP-heavy scratch space
distributed tools across compute clusters and teams
risks such as:
- accidental file deletion
- compute overload
- improper data sharing

Mitigation is discussed as iterative:

determine acceptable vs unacceptable risks during pilots
continuously harden security posture and reliability

Nvidia Siemens evaluation approach using “Nemo Gym”

Nemo Gym is positioned as a framework to evaluate and optimize agents (with evolution toward agent evaluation from runtime-style development).

Evaluation approach:

replay scenarios many times due to stochastic LLM behavior
move beyond “LLM-as-judge” by using existing analytical/ground-truth tools

Example workflow:

“Lighthouse” using Solido characterizer
foundries update PDKs regularly
agents help compare reference vs updated libraries
structured analytical tooling enables robust evaluation

Reported demonstrated improvements with Nemo Gym:

workflow reduced ~8 hours to ~2 hours
improved token efficiency (one mention: ~25% more token efficient)
higher success by debugging agent/tool interaction failures

Technical orchestration complexity (diagram discussion)

EDA agent orchestration is hard because:

too many tools / MCP servers plus semantic similarity can bloat context and confuse routing (e.g., “simulate” vs “analyze” descriptions)
stochastic behavior can lead to wrong tool paths
tools are compute-heavy, distributed across clusters, and cannot all be orchestrated directly from a single laptop/session

Siemens “Fuse Agent” orchestration solution:

a central gateway / unified MCP endpoint
described via a “USB-C dongle” analogy: connects distributed environments so the agent can request tasks across them

Agent “failure modes” and mitigation

Reported failure mode:

agents may attempt direct command-line calls or bypass intended tool pathways when skills/MCP descriptions are insufficient

Fix:

update skills files and MCP server descriptions based on Nemo Gym rollouts and failure analysis
goal: improve success rate and reduce wasted wall-clock time/tokens

Main speakers / sources

Tiffany Janzen / Tiff in Tech — host/founder
Amit Gupta — VP & GM, Siemens EDA
Tim Costa — VP & GM, Industrial Engineering, Nvidia

Deep-dive segment

Niranjan Nero Citipure — Siemens EDA, Central AI Product Manager
Josh Mayberry / Joshua Mabry — Nvidia, Senior Product Technical Lead

Repeatedly referenced products/orgs:

Siemens EDA: Fuse, Solido, Calibre OPC, Calibre-like parasitic extraction context
Nvidia: CUDA-X, Nemo, Nemotron, Open Shell, Nemo Gym, Agent Toolkit, MCP-related orchestration

Building the AI-native electronic design flow | Live from DAC with Tiff in Tech

Key takeaways