Cadence introduced an RTL Generation Agent for its ChipStack AI Super Agent that converts natural language specifications into production-ready RTL optimized for power, performance, and area. Early evaluations claim an average 24% area reduction and 18% power reduction versus pure foundation model code generation, with Honda R&D validating the agent on automotive SoCs ahead of Q4 2026 early access.
What is Covered in this Article
- RTL Generation Agent extending ChipStack from verification and debug into spec-to-RTL creation
- Vendor-claimed 24% area reduction, 18% power reduction, and 100% functional accuracy in early evaluations
- Honda R&D validation on safety-critical automotive SoCs for software-defined vehicles
- An RTL upgrade flow for adapting legacy RTL to new PPA and functional targets
- Competitive positioning against the Synopsys AgentEngineer L4 spec-to-RTL workflow
- Q4 2026 early-access availability across the ChipStack and InnoStack portfolio
The News: Cadence announced on September 22 a new RTL Generation Agent for its ChipStack AI Super Agent, extending the agentic platform from autonomous verification and debug into RTL creation and optimization. The agent automates power, performance, and area (PPA)-driven spec-to-RTL generation, RTL analysis, and refinement from natural language prompts, converting high-level specifications into production-ready RTL. In early evaluations, Cadence reports the agent delivered an average 24% area reduction and 18% power reduction versus pure foundation model code generation while producing 100% functionally accurate RTL. The release also introduces an RTL upgrade flow that adapts existing RTL to new architecture requirements, PPA targets, and functional requirements from high-level change descriptions. Honda R&D is evaluating the agent on advanced automotive SoCs, and the expanded ChipStack and InnoStack capabilities are expected to reach select early-access customers in the fourth quarter of 2026.
Cadence ChipStack Adds RTL Generation: Can Agents Code Production Silicon?
Analyst Take: Cadence is moving agentic AI up the design stack from checking silicon to writing it. The February 2026 ChipStack launch established autonomous verification and debug as the first commercial agentic workflow for front-end design, and the June introduction of a fully autonomous virtual engineer pushed verification autonomy further. RTL generation is a different order of ambition. Verification agents review work that human engineers produced. A generation agent produces the design itself, and every downstream implementation and signoff decision inherits its choices. Cadence’s differentiation is the pairing of LLMs with its implementation and signoff engines, so that generated RTL is optimized against real synthesis and power analysis rather than pattern-matched from training data. The 24% area and 18% power improvements are measured against pure foundation model code generation, and that baseline choice tells the strategic story: Cadence is arguing that raw LLMs writing Verilog will produce functionally plausible but physically wasteful silicon, and that the toolchain owner captures the value of closing that gap. The claim is consistent with Cadence’s technical leadership in AI integration. Its magnitude awaits production tapeouts.
Signoff-Grounded Generation Attacks the Weakness of Raw LLM Code
Foundation models have proven they can emit syntactically correct Verilog, and academic benchmarks such as VerilogEval have tracked steady gains, yet functional plausibility says nothing about physical quality. RTL that simulates correctly can still synthesize into bloated logic, miss timing, or burn power budget, and those failures surface months later in implementation. Cadence’s architecture routes generation through a hierarchy in which super agents orchestrate task-specific agents that call trusted EDA engines optimized for agentic workflows, closing the loop between what the model writes and what synthesis, power analysis, and signoff report back. Chin-Chi Teng, senior vice president and general manager in the Digital & Signoff Group at Cadence, said the approach pairs “agentic automation of spec-to-RTL and RTL refinement with our proven implementation and signoff engines” to deliver stronger correlation across the design flow. The moat in agentic RTL generation will belong to whoever owns the ground truth engines, since a model cannot optimize PPA it cannot measure. That framing favors the two EDA incumbents over foundation model vendors and internal hyperscaler research efforts such as NVIDIA’s VerilogCoder work, which generates code without production signoff feedback.
Honda Validation Puts Agentic RTL Inside Safety-Critical Silicon
The customer proof point is deliberately chosen. Honda R&D is evaluating the agent on advanced automotive SoCs, where ISO 26262 functional safety requirements, tight power and cost envelopes, and long qualification cycles make design teams among the most conservative buyers in the industry. Tomoya Nishino, chief engineer and general manager in Honda’s Digital Engine Development Division, said the company is working to improve productivity from specification through RTL development because the long development cycle of SoCs remains a major obstacle to software-defined vehicles. An automaker willing to attach its name to AI-generated RTL in safety-critical silicon signals that the productivity pressure has grown acute enough to overcome institutional caution.
The RTL upgrade flow may matter more to this buyer than a fresh generation. Automotive silicon carries decades of legacy RTL that must be adapted to new architecture requirements and PPA targets across model generations, and an agent that revises existing code while verifying functionality addresses the dominant workload of derivative design programs. Futurum has documented the same specification-to-silicon compression motive in Honda’s broader SDV compute strategy, where automotive OEMs are pulling chip development in-house on compressed timelines.
The Benchmark Baseline Flatters the Result
Cadence faces a high burden of proof that AI can produce adequate RTL, given industry skepticism, including indications that OpenAI’s chip design team was not able to accomplish it. A 24% area and 18% power advantage over pure foundation model code generation demonstrates that grounding generation in EDA engines beats not grounding it. The commercially decisive comparison is against RTL written by senior design engineers, and the release makes no claim there. The 100% functional accuracy figure comes from early evaluations of unstated scope, and functional accuracy on evaluation designs does not establish accuracy on novel microarchitectures where specification ambiguity, corner cases, and integration constraints dominate. Availability is also narrow. Early access for select customers begins in Q4 2026, so broad production evidence will not accumulate before 2027. Liability remains unresolved across the category. When an agent-generated block fails in a fielded vehicle, the allocation of responsibility among the EDA vendor, the foundation model provider, and the design team has no precedent, and automotive quality systems will demand an answer before agentic RTL ships in volume.
Cadence and Synopsys Are Now Contesting the Same Spec-to-RTL Ground
The competitive field converged on this exact capability within seven months. Synopsys unveiled its AgentEngineer-powered L4 agentic workflow in March 2026, generating RTL from natural language and formal specification, running lint checks, generating unit-level testbenches, and iterating verification to target objectives, with customers reporting 2x productivity gains and up to 5x in select cases. Synopsys then moved the workflow onto Microsoft Discovery in July with AMD as an active evaluator, adding cloud-scale compute and a marquee silicon partner.
Cadence’s counterposition is PPA depth. Synopsys frames its workflow around productivity and verification convergence, while Cadence claims quantified area and power outcomes tied to its Stratus, synthesis, and signoff engines. Cadence also brings portfolio breadth, with ChipStack, InnoStack, ViraStack, and the AuraStack packaging agent Futurum covered in July spanning digital, analog, verification, and system domains. The customer traction supports the strategy. Futurum’s analysis of Cadence’s Q2 FY 2026 earnings found revenue of $1.58 billion and a record $8.1 billion backlog with agentic AI cited as a demand driver, and Futurum’s CadenceLIVE 2026 coverage documented ChipStack production testing with NVIDIA, MediaTek, Google, Qualcomm, and Broadcom before feature expansion. That co-development discipline separates the EDA incumbents from the agent washing pervading enterprise software. It also means both vendors’ claims rest on lead customer evaluations that the broader market cannot yet inspect.
Read the complete announcement of the RTL Generation Agent on the Cadence newsroom.
What to Watch
- Whether early-access customers report area and power gains against human-written RTL rather than foundation model baselines
- Whether Honda advances from evaluation to production tapeout commitments on agent-generated automotive silicon
- Whether Q4 2026 early access converts to general availability and disclosed customer counts in 2027
- Whether Synopsys answers with PPA-quantified claims for its AgentEngineer spec-to-RTL workflow
- Whether Cadence discloses agentic AI revenue contribution or pricing structure in upcoming earnings calls
Sources
1. Cadence Expands ChipStack AI Super Agent with a New Agent for RTL Generation and Early PPA Optimization, Cadence, September 2026
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights from Futurum:
Rapidus’ IIM-1 Fab Completes Construction
Cadence AuraStack Extends Agentic Leadership to the Package
Cadence and Synopsys Accelerate Agentic EDA Race at Computex
Author Information
Brendan is Research Director, Semiconductors, Supply Chain, and Emerging Tech. He advises clients on strategic initiatives and leads the Futurum Semiconductors Practice. He is an experienced tech industry analyst who has guided tech leaders in identifying market opportunities spanning edge processors, generative AI applications, and hyperscale data centers.
Before joining Futurum, Brendan consulted with global AI leaders and served as a Senior Analyst in Emerging Technology Research at PitchBook. At PitchBook, he developed market intelligence tools for AI, highlighted by one of the industry’s most comprehensive AI semiconductor market landscapes encompassing both public and private companies. He has advised Fortune 100 tech giants, growth-stage innovators, global investors, and leading market research firms. Before PitchBook, he led research teams in tech investment banking and market research.
Brendan is based in Seattle, Washington. He has a Bachelor of Arts Degree from Amherst College.
