When my friend Simon Davidmann, whose framework anchored Part 1 of this series, laid out his agentic AI argument in recent months, one observation got less attention than his band-aid-versus-reset headline, yet it deserves more: The customers are doing it themselves. Hyperscalers and semiconductor houses are not waiting for vendors to ship finished AI products. They are building their own orchestration, their own agents, and their own models, on top of the simulation, formal, implementation, and signoff engines they already license. The EDA tools become engines; the intelligence layer above them increasingly belongs to the user.
As Simon pointed out, Samsung, Nvidia, and their peers are assembling internal AI stacks that treat vendor tools as engines beneath a domain-specific intelligence layer. But he attached a barbed caveat: Because that work stays behind corporate firewalls, it is âeffective and myopic at the same time.â Each giant rebuilds similar integrations in private; the lessons never propagate to the rest of the industryâand, worse for the rest of us, the giants get to learn from all their own data, something those outside the walls can only wish for. That is Davidmannâs Dilemma from Part 1 in its purest form: every player acting rationally and the ecosystem collectively worse off for it.
If you want evidence that DIY AI is now mainstream practice rather than hyperscaler exoticaâand, contra Simonâs caveat, that at least some of the lessons are starting to propagate in publicâthe DAC 2026 Engineering Track is the closest thing we have to proof. An honest qualifier belongs here: The most secretive giants still are not presentingâyou will not find Nvidia, Meta, or OpenAI walking through their internal design-AI stacks, and some of the industryâs most advanced users treat their flows as IP and stay home entirely. What the Track shows is the visible part of the iceberg. But the visible part has grown remarkably.
The Engineering Track is DACâs user conference within the conferenceâworking engineers presenting deployed results, peer-reviewed by other engineersâand this year, it reads like a census of do-it-yourself AI. What follows is my tour of the DIY AI wave, followed by its deliberate juxtaposition: users adopting the vendorsâ AI, with equally concrete numbers. Both are true simultaneously, and the tension between them is exactly where Simonâs framework earns its keep.
Exhibit A: Samsung teaches an RL agent to tune an SoC
Start with the presentation I would use to explain the entire trend to a skeptic: On Monday morning, Samsung Electronics engineers Hyungtae Park, Rira Oh, Sanggu Kim, Hyunjae Woo, and Youngsik Kim present âAdaptive QoS Optimization for SoC Performance Enhancement Using Dueling Double Deep Q-Networks.â The problem: Modern SoCs expose a thicket of quality-of-service knobsâarbitration, priorities, bandwidth allocations across many mastersâand manually tuning them across varying scenarios has become intractable. Samsungâs answer: Build an execution environment that runs the SoC pre-siliconâthink emulationâand set a reinforcement-learning (RL) agent loose in it. The agentâDeep Q-Networks with dueling and double-DQN enhancements, prioritized experience replay, the full modern RL toolkitâexplores the design space against reward functions built on throughput, latency, and power and discovers QoS settings that significantly outperform manual methods.
Notice the architecture of the achievementâand notice, per Part 1âs model-spectrum musing, that there is no LLM in sight: This is a user-trained optimization model, not a wrapped chatbot. The execution infrastructure that makes the exploration trustworthy comes from the EDA world; the intelligence that exploits it is Samsungâs own, purpose-built for Samsungâs architectures and Samsungâs performance targets for what they call âSystem Performance Verification.â That division of laborâvendor engines below, user intelligence aboveâis Simonâs observation rendered as a methodology paper, and it repeats across the track.
The DIY AI census
IBM turns agents loose on verification debugâand on tool development itself. In IBM Z hardware verification, agentic flows now automate failure triage and root-cause analysis, integrating design specifications, HDL, waveforms, coverage data, and prior issues through Model Context Protocol (MCP) servers, with customized agents extracting cone-of-influence logic and correlating transactionsâreporting an estimated 15% to 40% reduction in manual effort on one of the most expertise-hungry tasks in the flow.
A second IBM paper goes a level more meta: an agentic framework that generates EDA utilities themselves from user specifications using reusable MCP building blocks, reducing one structural-verification toolâs development from an estimated four person-weeks to under 30 minutes. Note MCP appearing twice: The connective-tissue standard from Part 1âs data-and-glue layer (where Perforce pairs it with RAG for design-data context) is showing up in user methodology papers barely a year after entering the conversation. The connective tissue is arriving. And it raises a question Simon put to me directly, worth carrying into Tuesdayâs Build-versus-Buy panel: Why are the big EDA vendors not shipping these utilities themselvesâand why are users such as IBM building rather than buying from the startups next door?
An on-premises LLM babysits emulation regressions
A flagship-SoC team from Samsung presents an unattended software-testing framework for pre-silicon emulation that pairs fast-boot virtualization with an on-premises LLM that reads kernel panics and decides: genuine software bug or non-critical bring-up artifact? The detail to savor is âon-premisesââthe confidentiality requirements of semiconductor design mean the userâs DIY AI runs inside the userâs walls, on top of the vendorâs emulation platform. Another team presents an on-premises, multi-agent âTeam of Expertsâ pipelineâa classifier agent routing to specialized sub-agents for UVM code generation, log parsing, and specification retrieval across JEDEC, LPDDR, and HBM documentsâbuilt with RAG and finetuning precisely so that no design data leaves the building.
The knowledge layer gets formalized
A Samsung team presents an ontology-driven knowledge graph that transforms hundreds of pages of interlinked design documents into a queryable graph feeding a Graph-RAG chatbotâthe specification-engineering thread from Part 1âs front-end stage, implemented by users for users. In the POWER processor verification environment, ML-driven analysis correlates test components with coverage events and composes âsuper-testsâ that make up 86% of recommended test cases for hard-to-hit scenarios.
Implementation and analog join in
Texas Instruments describes a deep-RL analog sizing toolâwith custom reward shaping, optimizing across all PVT and Monte Carlo cornersâdeployed at scale with over 100 circuits optimized to signoff quality. An automotive SoC team from Renesas presents a Random Forest framework predicting die size from design specifications with R2 up to 0.95 across 1,800 RTL modules. On the Gladiator stage, a team from Samsung uses LLMs to interpret netlists and cluster DRAM peripheral cells for placementâexpert structural knowledge written as natural-language prompts rather than rule-based codeâcutting design-rule violations by 80.6%, while Samsungâs Sungsu Byun applies gradient-boosted models to switch-cell optimization against RedHawk-reported voltage variation. A neuro-symbolic framework presented by Cadence called Auto-Chk compiles natural-language signoff intent into deterministic checkersâ120+ checklist items in production, 20Ã faster checker development. Even formal verification gets the treatment: AI-generated âhelperâ invariants converting bounded proofs into full closure. And one paper turns AI on the infrastructure itself, using graph neural networks and RL to forecast EDA compute needs, cutting cloud overprovisioning by 60% to 80%.
Add it up, and the pattern is unmistakable: RL agents, on-prem LLMs, knowledge graphs, multi-agent pipelines, neuro-symbolic compilersâbuilt by users, running on vendor engines, presented with deployment numbers rather than aspirations. Hold each one up to Davidmannâs Test, and the census gets even more interesting: Most of this speeds up what their teams already verifiedâdramatically, honestly, and with numbers. A fewâthe invariants that convert bounded proofs into full closure, the QoS exploration no human team could have sweptâedge toward changing what can be verified at all.
The juxtaposition: vendor AI, adopted and audited
Now letâs look at the other half of the picture, because the same Track that documents the DIY AI wave also documents users adopting vendor AIâand, in the best Engineering Track tradition, auditing it in public.
A user methodology paper on Cadence Voltus InsightAI examines AI-driven power-grid reinforcement for IR closureâand does something refreshingly honest: It presents a detailed design-impact analysis for working with the toolâs capabilities and around its limitations. That sentence is the Engineering Trackâs entire value proposition in miniature. Broadcomâs APD AI team and Synopsysâs RedHawk-ET team jointly present an early 3DIC thermal flow, correlated to silicon within 3°C (Part 3 returns to it)âvendor AI and user expertise co-developed rather than bought off the shelf. Intel presents its library QA framework built on Siemensâs Solido Crosscheck, native checks fine-tuned alongside custom ones. An AI-driven thermal-aware placement methodology explores hundreds of constrained scenarios for 3D chiplet stacks. And in the special sessions, Sandiskâs Mariana Shenker shows production AI workflows for specification-driven developmentâthe user-experience counterpart to the âSpec-Tacular!â sessionâs startup pitches.
So which is it: users building their own or users adopting the vendorsâ? The Trackâs answer is both, sorted by where the intelligence needs to live. Where the problem is deeply proprietaryâan SoCâs QoS behavior, a companyâs debug tribal knowledge, confidential failure logsâthe AI is homegrown and on-premises, with vendor engines supplying the physics and the simulation truth underneath. Where the problem is common to everyoneâIR closure, library QA, thermal analysisâvendor AI embedded in the platforms wins on economics. The boundary between those two zones is, I suspect, where the next five years of EDA business models get decided. Simonâs holistic reset, if it comes, will have to serve both sides of that line; Mondayâs DAC Pavilion panel âAgentic AI in EDA: Whoâs in Control?â and Tuesdayâs âBuild vs Buy: Who Owns the Intelligence Behind Tomorrowâs Chips?â put that exact question on stage, and Simon himself joins Wednesdayâs exhibitor forum panel, âHarnessing AI for SoC Verification: Disruptive or Collaborative?ââthe perfect closing argument for this articleâs thesis, argued live.
For the startup-side complement to these user stories, Mondayâs Veriest-hosted exhibitor forum session, âBeyond the Hype: Innovation in AI for Chip Design and Verification,â showcases what the agentic layer from Part 1âs landscape is building for exactly these users.
The itinerary, complete
Three parts, one arc: Part 1, the landscape of who; Part 2, the users quietly building their own intelligence on the industryâs engines; and Part 3, the sessions on creating AI chipsâthe end product this whole industry exists for.
And if Part 1âs Expedia comparison did not convince you, here is a culinary one: Think of DAC Explorer as the conferenceâs AI sous-chef. Hand it the ingredients you fancy: say, a base of RL, some on-premises LLMs, and a pinch of MCP, and it finds the recipeâa personalized schedule of sessions, plated and ready to export. No reservations are required, although seats at the popular sessions do go fast.
Registration is at dac.com, and every session mentioned here is in the official DAC 2026 program. See you in Long Beach.
Frank Schirrmeister writes about semiconductors, EDA, and system design. He serves as program chair of the DAC Engineering Track; this series is written in that capacity, not on behalf of his employer, Synopsys. Opinions are his own.
Read also:
Strategies for Addressing More Complex Custom Chip Design
How Chip Startups Are Changing the Way Chips Are Designed
AI Chips Shifting from Round to RectangularÂ
Leave a Reply
You must Register or Login to post a comment.