WP_Term Object
(
    [term_id] => 25181
    [name] => Undo
    [slug] => undo
    [term_group] => 0
    [term_taxonomy_id] => 25181
    [taxonomy] => category
    [description] => 
    [parent] => 157
    [count] => 8
    [filter] => raw
    [cat_ID] => 25181
    [category_count] => 8
    [category_description] => 
    [cat_name] => Undo
    [category_nicename] => undo
    [category_parent] => 157
)
            
AI Pre Silicon Modeling Webinar
WP_Term Object
(
    [term_id] => 25181
    [name] => Undo
    [slug] => undo
    [term_group] => 0
    [term_taxonomy_id] => 25181
    [taxonomy] => category
    [description] => 
    [parent] => 157
    [count] => 8
    [filter] => raw
    [cat_ID] => 25181
    [category_count] => 8
    [category_description] => 
    [cat_name] => Undo
    [category_nicename] => undo
    [category_parent] => 157
)

Is AI Ready to Help Chip Teams Get the Architecture Right?

Is AI Ready to Help Chip Teams Get the Architecture Right?
by Admin on 11-18-2026 at 8:00 am

Key takeaways ▼
AI and precision in chip architecture
By Michael Sanie

Every silicon executive worries about schedule. A slipped tape-out is painful. A chip that works but loses to the previous generation is worse. The question that matters most is simple. Will this architecture win on the workloads customers will actually run?

The best teams answer that question long before RTL. The leading CPU, GPU and AI accelerator companies invest as much in architectural and behavioral modeling as they do in implementation and verification. They run real software on models of silicon that does not exist yet, and they tune cache sizes, interconnect bandwidth, memory hierarchies and scheduling policies against what those models show.

Register Here

The Quiet Advantage

That discipline is one of the least discussed reasons the leaders stay ahead. Building model infrastructure, correlating it to silicon and keeping it alive as the design changes takes years and large, dedicated teams. Few companies can afford it, and fewer can sustain it across generations.

Generative AI puts that advantage in play. AI can already write model code, generate tests and help engineers move through codebases with millions of lines. The open question is whether AI widens the gap for teams that already own the models, the workloads and the data, or narrows it for everyone else.

Opportunity and Risk

AI opens real opportunities in modeling. It can build models faster, bring more workloads up sooner, and help keep models correlated as the design changes. With physics-informed models, AI may even enable teams to weigh some PPA tradeoffs during architectural exploration, well before implementation.

It also adds risk. In pre-silicon modeling, correct means precise, deterministic and correlated with a design that is still moving. Firmware, drivers, libraries and applications run on top, often across architectural models, simulation and emulation at once. Every layer is a place for something to break. When it does, workloads stop running and decisions rest on less data than they deserve. AI also makes mistakes that look plausible. A model that is plausible but wrong is worse than no model, because teams trust it.

The Trade-Offs

As AI writes more of the model and more of the tests, the hard question shifts from whether the code is correct to what actually happened. Teams need to know what the AI changed, why it changed it, and how the software behaved on the model when a result went wrong.

That is the trade-off. AI shortens the time to build a model, but it can lengthen the time to understand one. Engineers who once wrote every line now inherit code they did not write, running software stacks they did not build. When a workload fails or a result looks off, someone still has to explain why. Teams that keep the context behind each change, workload and validation run will move faster with AI. Teams that do not will spend the time they saved explaining results they cannot trust.

The Panel

I’m looking forward to moderating the panel with leaders from across the pre-silicon modeling, validation, and software  stack

(listed alphabetically by company):

● Alex Starr, Corporate Fellow, AMD
● John Rayfield, SVP, Advanced Computing Lab, Samsung
● Frank Schirrmeister, Executive Director, Synopsys
● Greg Law, Co-founder and CEO, Undo

Between them, they cover silicon design at scale, EDA and virtual prototyping, and the tools engineers use to understand complex software running on these models. They do not all see this problem the same way, and the discussion will not ask them to.

The debate will focus on five questions:

1. Where is AI already changing how architectural models are built, and where is it still hype?
2. Does AI make pre-silicon models and validation easier to trust, or harder?
3. As AI does more of the work, what context about workloads, changes and execution must teams capture to stay in control?
4. When AI writes the model, how do you know it is right, and how fast can you find out when it is not?
5. What should engineering leaders change in their teams and flows in the next twelve months?

The teams that win the next generation will know where AI can be trusted with their architecture, and where it cannot. If you build CPUs, GPUs or AI accelerators, this conversation is for you.

Is AI Ready to Transform Pre-Silicon Modeling?
SemiWiki virtual panel: Thursday, December 10, 2026
10:00 am PT / 1:00 pm ET / 6:00 pm GMT
Hosted by Daniel Nenni, SemiWiki, moderated by Michael Sanie

Register Here

AI Pre Silicon Modeling Webinar

Also Read:

DAC 2026: How Undo Makes AI for Chip Design Work Better

The Other Side of Bug Localization

Enabling AI to Understand Complex Runtime Behavior for Accurate, Automated Root Cause Analysis — DAC 2026

Share this post via:

Comments

There are no comments yet.

You must register or log in to view/post comments.