# Reinvent Semiconductor Design With Microsoft Discovery and Synopsys Agent Engineers (Presented by Microsoft)

## Abstract

Semiconductor teams face exploding design complexity, shrinking windows, and increasing verification challenges. This session shows how the Microsoft Discovery AI for Science platform, combined with Synopsys Agent Engineers, introduces an agentic approach to electronic design automation (EDA) that automates routine steps, accelerates expert decision‑making, and keeps data secure on Azure. Discovery’s Science Copilot orchestrates multi‑agent plans leveraging Azure AI resources and a graph‑grounded knowledge layer, while Synopsys agents drive proven flows levering domain expertise—spanning spec‑to‑RTL, simulation, timing, and sign‑off—directly on Azure HPC infrastructure. See how novel methodology integrates with existing Synopsys tools, and how NVIDIA‑accelerated Azure AI infrastructure results in faster time‑to‑PPA, fewer iterations, and more predictable schedules.

## AI Summary

- Introduced Microsoft Discovery, an enterprise agentic platform designed for next-generation R&D work, focusing on silicon design, chemistry, material science, healthcare, and physics.
- Highlighted the Discovery Engine, which decomposes complex tasks into executable workflows, using a task planner and agents to optimize outcomes.
- Demonstrated a spec-to-RTL workflow for chip design, where the platform generated and optimized Verilog code and test benches, significantly improving the design process.
- Showed a real-world example of discovering a new battery electrolyte that reduced lithium use by 70%, and another example of creating a novel, environmentally friendly immersion coolant for data centers.
- Presented a case where the platform optimized a cooling fan design for Microsoft Surface devices, balancing noise reduction and thermal efficiency.
- Emphasized the potential for generative AI to produce innovative, nature-inspired designs, such as an enhanced heat sink that outperformed human designs in heat dissipation and weight.

## Transcript

**[00:00:08 – 00:00:10]** Thank you so much. Hello, everyone.

**[00:00:10 – 00:00:13]** Thank you for making it to the session.

**[00:00:14 – 00:00:17]** Today, I'm going to talk about how this newly introduced

**[00:00:17 – 00:00:22]** platform from Microsoft Agentech platform combined with the

**[00:00:22 – 00:00:26]** power of Synopsys tools is actually reinventing silicon design.

**[00:00:26 – 00:00:28]** Big focus is going to be silicon design, but also I'm going

**[00:00:28 – 00:00:32]** to cover some examples in other scientific areas like

**[00:00:32 – 00:00:37]** chemistry, material science, and health care life sciences as well.

**[00:00:37 – 00:00:40]** But before we get into that, I want to actually cover a little

**[00:00:40 – 00:00:43]** bit details about the platform, what that is, and what we've

**[00:00:43 – 00:00:46]** been able to accomplish so far.

**[00:00:46 – 00:00:49]** So we're in an interesting time today, right?

**[00:00:49 – 00:00:53]** I mean, we have all of these emerging technologies, stuff that

**[00:00:53 – 00:00:57]** you're hearing in the news, you're also hearing at this event as well.

**[00:00:57 – 00:01:01]** There's new advancements in AI, of course, with the deep

**[00:01:01 – 00:01:03]** reasoning and everything coming on.

**[00:01:03 – 00:01:06]** The level of complexity of the problem that can actually

**[00:01:06 – 00:01:11]** be solved with AI is advancing on an hourly basis.

**[00:01:11 – 00:01:15]** HPC innovation on the hardware side, all the way from GPUs,

**[00:01:15 – 00:01:19]** CPUs, but then also looking at quantum, which is actually

**[00:01:19 – 00:01:20]** becoming almost real.

**[00:01:20 – 00:01:23]** I mean, it's supposed to be something that was decades

**[00:01:23 – 00:01:25]** away, but that's not true anymore.

**[00:01:25 – 00:01:28]** We are actually looking at some real-world problems

**[00:01:28 – 00:01:31]** being solved by quantum, and that's happening today.

**[00:01:31 – 00:01:35]** And of course, we've got robotics, that's a new emerging area,

**[00:01:35 – 00:01:40]** and these are basically embodied AIs that are literally walking

**[00:01:40 – 00:01:42]** out of the data center.

**[00:01:42 – 00:01:46]** So when we look at this more offensive as an opportunity

**[00:01:46 – 00:01:49]** to see how we can leverage these technologies in the

**[00:01:49 – 00:01:54]** scientific research and engineering innovation, we think of this

**[00:01:54 – 00:01:59]** agents that we keep hearing about and humans combined now have

**[00:01:59 – 00:02:04]** the power to start from scientific reasoning all the way to execution

**[00:02:04 – 00:02:08]** to actually experimentation analysis in a continuous research

**[00:02:08 – 00:02:11]** loop, Which is basically leveraging all of the technologies that

**[00:02:11 – 00:02:14]** you see at the bottom.

**[00:02:15 – 00:02:22]** Now, we, of course, want to make sure that we are able to provide

**[00:02:22 – 00:02:26]** access to all of these technologies to the modern R&D teams to

**[00:02:26 – 00:02:29]** be able to amplify their ingenuity.

**[00:02:29 – 00:02:33]** But there are challenges because these technologies that are off

**[00:02:33 – 00:02:38]** the shelf are not necessarily ready to be consumed by the R&D teams.

**[00:02:39 – 00:02:43]** You've got, of course, every modern RNG team, they use

**[00:02:43 – 00:02:47]** a bunch of engineering tools, they have 10 plus tools, they have

**[00:02:47 – 00:02:52]** complex workflows, and usually when you start to pick a solution like

**[00:02:52 – 00:02:56]** this, it is optimized for their own specific use case, it has their

**[00:02:56 – 00:03:00]** own infrastructure requirement, it has its own architectural

**[00:03:00 – 00:03:04]** limitations, which causes a lot of these solutions to become bespoke.

**[00:03:04 – 00:03:06]** They become very, very hard to scale, and

**[00:03:06 – 00:03:09]** extremely hard to secure.

**[00:03:09 – 00:03:12]** And then this whole relentless pace of innovation and pace

**[00:03:12 – 00:03:17]** of introduction of new technology is keeping all of these R&D teams

**[00:03:17 – 00:03:19]** in a constant benchmarking state.

**[00:03:19 – 00:03:21]** They're always looking at new technologies that are

**[00:03:21 – 00:03:26]** coming out, new architectures, new orchestration frameworks,

**[00:03:26 – 00:03:28]** new protocols even.

**[00:03:28 – 00:03:30]** And they all become very interesting because you

**[00:03:30 – 00:03:32]** can actually relate to them, you feel like, hey, I can

**[00:03:32 – 00:03:33]** use this in my day-to-day work.

**[00:03:34 – 00:03:36]** But when you start to really adopt this technology, you

**[00:03:36 – 00:03:41]** realize with these infrastructure moving so quickly, you do

**[00:03:41 – 00:03:47]** not have a stable framework, a stable platform to build upon.

**[00:03:47 – 00:03:50]** And then, of course, you have the trust and compliance and

**[00:03:50 – 00:03:54]** security, which are top of mind for the entire engineering community.

**[00:03:54 – 00:03:58]** It's always been the case, it is table stakes, which is

**[00:03:58 – 00:04:04]** not necessarily guaranteed by all the new up-and-coming technologies.

**[00:04:05 – 00:04:08]** So this is the reason, like if you really wanted to solve

**[00:04:08 – 00:04:11]** some of those problems, you really needed to think out

**[00:04:11 – 00:04:14]** of the box and bring an innovative solution to the market.

**[00:04:15 – 00:04:17]** That's why we introduced Microsoft Discovery.

**[00:04:17 – 00:04:21]** It's an enterprise agentic platform that is designed

**[00:04:21 – 00:04:24]** for the next generation R&D work.

**[00:04:24 – 00:04:29]** And it's really combining the power of hardware infrastructure,

**[00:04:29 – 00:04:33]** orchestration, agent tech orchestration, along with

**[00:04:33 – 00:04:37]** cognition that can provide a natural language interface to being

**[00:04:37 – 00:04:39]** able to solve complex workflows.

**[00:04:39 – 00:04:42]** So the three pillars

**[00:04:42 – 00:04:45]** For the service, first one is the Discovery Engine.

**[00:04:45 – 00:04:47]** I'm going to talk more about this.

**[00:04:47 – 00:04:49]** There is a comprehensive R&D enterprise platform, what

**[00:04:49 – 00:04:51]** we call as the enterprise grade.

**[00:04:51 – 00:04:54]** Now, this is a big piece, even though it sounds like,

**[00:04:54 – 00:04:56]** hey, it's a platform.

**[00:04:56 – 00:04:59]** But what makes it really enterprise grade is what

**[00:04:59 – 00:05:00]** I'm going to talk about.

**[00:05:00 – 00:05:03]** And then, of course, trust and security, and what does that

**[00:05:03 – 00:05:07]** mean in the context of R&D work.

**[00:05:08 – 00:05:11]** So let me talk about Discovery Engine a little bit.

**[00:05:11 – 00:05:14]** Maybe you all know, I mean, this whole pace of development

**[00:05:14 – 00:05:18]** over the last couple of years has been blinding, right?

**[00:05:18 – 00:05:21]** Something like this, a chart like this, typically, historically,

**[00:05:21 – 00:05:24]** you would have spanned multiple years, sometimes even decades,

**[00:05:24 – 00:05:26]** to talk about technology evolution.

**[00:05:26 – 00:05:29]** But nowadays, all of that timeframe has shrunk, and

**[00:05:30 – 00:05:31]** we're looking at months.

**[00:05:31 – 00:05:33]** for the evolution to happen.

**[00:05:33 – 00:05:37]** So sometime last year, the agents obviously emerged and

**[00:05:37 – 00:05:41]** the adoption started in a very simplistic scenario of

**[00:05:41 – 00:05:43]** single user, single agent.

**[00:05:43 – 00:05:46]** And we quickly evolved into agents now collaborating with

**[00:05:46 – 00:05:50]** each other and now able to execute a complex workflow.

**[00:05:50 – 00:05:54]** So that same user now can actually have a long-running workflow.

**[00:05:54 – 00:05:57]** They can actually execute complex tasks.

**[00:05:57 – 00:06:01]** To now, where you're actually talking about multiple users

**[00:06:01 – 00:06:05]** and teams of users and teams of agents are collaborating with

**[00:06:05 – 00:06:10]** each other to actually accomplish an even more complex task.

**[00:06:10 – 00:06:13]** So we wanted to see how do we bring this kind of

**[00:06:13 – 00:06:18]** thinking to engineering work, to research work, right?

**[00:06:18 – 00:06:20]** So that's why we built the discovery engine.

**[00:06:20 – 00:06:22]** So this is really the key.

**[00:06:22 – 00:06:27]** This is the cognition or cognitive brain behind discovery.

**[00:06:27 – 00:06:31]** So at the top, you have the task planner, where you have

**[00:06:31 – 00:06:33]** an intent that the user provides.

**[00:06:33 – 00:06:38]** It actually decomposes it into independently executable tasks.

**[00:06:38 – 00:06:42]** And these are the tasks that are then attached to a specific agent.

**[00:06:42 – 00:06:44]** But you know what task needs to be accomplished.

**[00:06:45 – 00:06:49]** It knows what agents are available for it to use, attaches agents

**[00:06:49 – 00:06:51]** to those specific tasks.

**[00:06:51 – 00:06:54]** Agents have a much broader view of the tools and the models

**[00:06:55 – 00:06:57]** and the data sets it has available.

**[00:06:57 – 00:06:59]** And so be able to configure this in the right fashion

**[00:06:59 – 00:07:02]** to be able to accomplish that task.

**[00:07:02 – 00:07:06]** A lot of this happens as part of the task decomposition

**[00:07:06 – 00:07:08]** and creating a plan for execution.

**[00:07:08 – 00:07:11]** But we all know, whenever we execute these plans, they

**[00:07:11 – 00:07:15]** are rarely unidirectional, right?

**[00:07:15 – 00:07:16]** You always branch out.

**[00:07:16 – 00:07:17]** You have certain decision-making.

**[00:07:17 – 00:07:21]** You've got to experiment with multiple things at the same time.

**[00:07:21 – 00:07:25]** You may hit dead ends on some of the experiments.

**[00:07:25 – 00:07:28]** In some of them, you are not getting convergence, and you're

**[00:07:28 – 00:07:32]** having to iterate over the same thing again and again.

**[00:07:32 – 00:07:36]** So you need a system really that can now be learning in

**[00:07:36 – 00:07:39]** the loop while you're doing all these experiments, not just from

**[00:07:39 – 00:07:43]** the intermediate data it produces, but also the feedback that the

**[00:07:43 – 00:07:49]** human's giving of the misdirections or failures or iterations that

**[00:07:49 – 00:07:51]** it experiences all the way through.

**[00:07:51 – 00:07:54]** So this is what Discovery Engine really accomplishes

**[00:07:54 – 00:08:00]** by having this in a context graph that can now annotate all of

**[00:08:00 – 00:08:05]** this information and be able to use this to refine the task execution

**[00:08:05 – 00:08:07]** and optimize the outcomes.

**[00:08:07 – 00:08:09]** And why is this important, right?

**[00:08:09 – 00:08:10]** Especially for engineering workflows that are hugely

**[00:08:11 – 00:08:12]** complex, why is this important?

**[00:08:12 – 00:08:15]** It's important because the better insights that you can

**[00:08:15 – 00:08:19]** generate from the data and the better decisions you can make,

**[00:08:19 – 00:08:22]** The faster your workflows are going to be, the more

**[00:08:22 – 00:08:24]** efficient the workflows are going to be, and the better

**[00:08:24 – 00:08:28]** your quality of outcome will be.

**[00:08:29 – 00:08:32]** The comprehensive R&D enterprise program, I will talk about

**[00:08:32 – 00:08:32]** this a little bit.

**[00:08:32 – 00:08:34]** I think this is a very important topic.

**[00:08:34 – 00:08:38]** So this is the platform view for discovery.

**[00:08:38 – 00:08:42]** So we're focusing on a few industry verticals, right?

**[00:08:42 – 00:08:45]** So there's a chemistry and material science.

**[00:08:45 – 00:08:49]** There's healthcare, life sciences, new drug discovery.

**[00:08:49 – 00:08:51]** And then physics, on the physics, we have physics

**[00:08:51 – 00:08:54]** and semiconductors, multi-physics, system design, et cetera.

**[00:08:54 – 00:08:57]** These are the verticals that we are currently focused on,

**[00:08:57 – 00:08:59]** and we're working with our industry-leading partners to try

**[00:08:59 – 00:09:03]** and offer solutions in this space.

**[00:09:03 – 00:09:06]** In Semiconductor in particular, we have our partner synopsis.

**[00:09:06 – 00:09:09]** We see that's part of the title of this session.

**[00:09:09 – 00:09:11]** And we have team members sitting here in the audience as well.

**[00:09:11 – 00:09:13]** So thank you again for participation.

**[00:09:13 – 00:09:18]** And then of course, the platform capability is horizontal.

**[00:09:18 – 00:09:22]** In terms of agentic orchestration and infrastructure orchestration,

**[00:09:22 – 00:09:26]** cognition ability to be able to convert a complex problem

**[00:09:26 – 00:09:29]** into individually executable tasks.

**[00:09:29 – 00:09:33]** So expanding this into other verticals is definitely something

**[00:09:33 – 00:09:35]** that's on our roadmap as well.

**[00:09:36 – 00:09:39]** And we are building on the solid foundation of Azure, right?

**[00:09:39 – 00:09:43]** So from an AI perspective, of course, we have the AI Foundry

**[00:09:43 – 00:09:45]** that we're building on top of.

**[00:09:45 – 00:09:49]** We've got the actual hardware, a scalable hardware as part of Azure.

**[00:09:49 – 00:09:54]** HPC, of course, is huge when it comes to engineering workflows,

**[00:09:54 – 00:09:57]** specifically chip design.

**[00:09:57 – 00:10:02]** All the way from CPUs and GPUs, and in this case, when you're

**[00:10:02 – 00:10:06]** looking at generative AI, we have heavy use of GPUs, and so looking

**[00:10:06 – 00:10:12]** at NV and NC series, which use the A100s, H100s, and also latest

**[00:10:12 – 00:10:17]** Blackwell GPUs as well, and then of course all the data services,

**[00:10:17 – 00:10:20]** but the middle layer that you see is really the discovery, right,

**[00:10:20 – 00:10:25]** that's the discovery service, that's the discovery platform.

**[00:10:25 – 00:10:27]** Discovery Engine I just talked about.

**[00:10:27 – 00:10:29]** The Science Bookshelf and Science Supercomputer I'm

**[00:10:29 – 00:10:31]** gonna dig further into.

**[00:10:31 – 00:10:35]** In a nutshell, Science Bookshelf is just the mechanism of

**[00:10:35 – 00:10:38]** making sense of your own data, generating insights,

**[00:10:38 – 00:10:41]** generating knowledge out of data.

**[00:10:41 – 00:10:45]** And then Science Supercomputer is really kind of like a collaboration

**[00:10:45 – 00:10:49]** environment for multiple users to come and interact with agents

**[00:10:49 – 00:10:55]** in a governance-guided, role-based access control-driven environment.

**[00:10:55 – 00:10:58]** I'm going to talk more about both of these.

**[00:10:58 – 00:11:02]** So Science Bookshelf is, like I was saying, is just a way

**[00:11:02 – 00:11:03]** to make sense of your data.

**[00:11:03 – 00:11:05]** Now, data could be coming from various sources.

**[00:11:05 – 00:11:09]** You could have open source data available on the internet.

**[00:11:09 – 00:11:12]** You could have paywall data coming through subscriptions, et cetera.

**[00:11:12 – 00:11:15]** And then you have enterprise data, your own data, your

**[00:11:15 – 00:11:19]** own prior designs, your own prior experiments that you may have done.

**[00:11:19 – 00:11:25]** All of that can be brought into the context of your specific workflows.

**[00:11:26 – 00:11:29]** So, Bookshelf really, it's the core founding sort of

**[00:11:30 – 00:11:35]** technology for Bookshelf is this technology called GraphRAG,

**[00:11:35 – 00:11:40]** which is a modification of the existing VectorRAG technology,

**[00:11:41 – 00:11:44]** and this GraphRAG really enables the Bookshelf to start

**[00:11:44 – 00:11:48]** generating insights and really extract, you know, synthesize

**[00:11:48 – 00:11:52]** the information, synthesize the knowledge that may be

**[00:11:52 – 00:11:56]** Distributed across multiple data sets, multiple files,

**[00:11:56 – 00:11:58]** multiple documents, right?

**[00:11:58 – 00:12:00]** This is what really gives it the intelligence to do

**[00:12:00 – 00:12:05]** multi-hop reasoning and be able to generate that insight.

**[00:12:08 – 00:12:11]** And science supercomputer, like I was saying, is basically

**[00:12:11 – 00:12:13]** kind of like HPC in a box.

**[00:12:13 – 00:12:17]** It's like a collaboration environment where multiple

**[00:12:17 – 00:12:19]** users can come in, be onboarded.

**[00:12:19 – 00:12:22]** They act and behave like it's a team.

**[00:12:22 – 00:12:24]** They can interact with the agents.

**[00:12:24 – 00:12:26]** They have data assets that they are accessing.

**[00:12:26 – 00:12:28]** They've got workflows, they have got tool sets that are all

**[00:12:28 – 00:12:33]** enabled within that environment, as well as the infrastructure

**[00:12:33 – 00:12:36]** that's now orchestrated, which is specific to the workflow.

**[00:12:36 – 00:12:37]** So now,

**[00:12:37 – 00:12:43]** You imagine you start with a task, and your task has

**[00:12:43 – 00:12:44]** been attached to the agent.

**[00:12:44 – 00:12:47]** It basically understands all the agents that are available

**[00:12:47 – 00:12:50]** to it, and it can make the selection from the models

**[00:12:50 – 00:12:55]** and tools and agents, and then be able to execute that in the

**[00:12:55 – 00:12:58]** supercomputer instance, which is appropriately sized, because it's

**[00:12:58 – 00:13:01]** extremely important to not have this very expensive infrastructure

**[00:13:01 – 00:13:03]** to be running all the time.

**[00:13:03 – 00:13:05]** So you want infrastructure, cost level controls, performance

**[00:13:05 – 00:13:11]** level controls, they are all built in as part of the supercomputer.

**[00:13:13 – 00:13:17]** And then of course, the last pillar here is the trust and security.

**[00:13:17 – 00:13:19]** This is huge. This is basically foundational.

**[00:13:19 – 00:13:23]** This is table stakes for every service that Microsoft builds.

**[00:13:23 – 00:13:27]** And we are building on top of Azure, so we follow all

**[00:13:27 – 00:13:30]** the security protocols and security implementation

**[00:13:30 – 00:13:33]** policies that Azure has.

**[00:13:33 – 00:13:37]** And from a trust perspective, we are basically looking at

**[00:13:37 – 00:13:40]** human in the loop for the entire process, right, through

**[00:13:40 – 00:13:42]** the entire discovery process.

**[00:13:42 – 00:13:45]** And then of course, with all the investments that Microsoft

**[00:13:45 – 00:13:49]** is making on security, as well as new next-generation

**[00:13:49 – 00:13:53]** technology, it's also future-proof.

**[00:13:54 – 00:13:56]** So let me talk about this a little bit.

**[00:13:56 – 00:14:01]** So I mentioned scientific method earlier a little bit.

**[00:14:01 – 00:14:03]** The traditional scientific method, we've all learned

**[00:14:03 – 00:14:06]** this in primary school, it starts with the knowledge.

**[00:14:06 – 00:14:10]** You generate a hypothesis, and based on the hypothesis,

**[00:14:10 – 00:14:12]** you configure an experiment.

**[00:14:12 – 00:14:15]** You run the experiment, produce data, you analyze the results,

**[00:14:15 – 00:14:18]** take the learning, and go iterate.

**[00:14:18 – 00:14:24]** So we believe that the engineering workflows that we have today,

**[00:14:24 – 00:14:27]** they actually, the scientific, the principles of this iterative

**[00:14:27 – 00:14:30]** scientific method applies to those as well, right?

**[00:14:30 – 00:14:33]** So you have to have the knowledge, knowledge in terms of, for

**[00:14:33 – 00:14:35]** engineering workflow for chip design, for example, your

**[00:14:35 – 00:14:39]** own specification data, your past designs, your IP information,

**[00:14:39 – 00:14:41]** user guides, etc.

**[00:14:41 – 00:14:43]** Feeds into a hypothesis which actually comes up

**[00:14:43 – 00:14:45]** with an experiment that you've got to design, which

**[00:14:46 – 00:14:48]** have an appropriate selection, appropriate configuration

**[00:14:48 – 00:14:51]** of tools, and appropriate selection and configuration

**[00:14:51 – 00:14:54]** of infrastructure, running those jobs in a reliable fashion, be able

**[00:14:54 – 00:14:56]** to extract learnings, et cetera.

**[00:14:56 – 00:15:02]** So at the core of engineering workflows, we see the same

**[00:15:02 – 00:15:05]** scientific method being applied.

**[00:15:05 – 00:15:11]** And how do we think about this in the context of chip design, right?

**[00:15:11 – 00:15:13]** This is a bit of an eye chart, but the folks who

**[00:15:13 – 00:15:16]** understand chip design, they'll probably relate to this.

**[00:15:16 – 00:15:20]** You know, starting from the left, intent to spec, to spec

**[00:15:20 – 00:15:24]** to RTL, then RTL to layout, that's kind of how it is laid out.

**[00:15:24 – 00:15:29]** But the idea here is, we think of these, each of the boxes

**[00:15:29 – 00:15:32]** that you see over here are individual agents.

**[00:15:32 – 00:15:35]** And in this case, we're using, you know, these agents are coming from

**[00:15:35 – 00:15:40]** Synopsys that are hyper-optimized into their own specific domain.

**[00:15:40 – 00:15:41]** They understand their data well.

**[00:15:41 – 00:15:45]** They're grounded in their own domain data through GraphRAG.

**[00:15:45 – 00:15:51]** And we think of this workflow and these agents as composable.

**[00:15:51 – 00:15:54]** And they are, as they are moving through the workflow,

**[00:15:54 – 00:15:58]** acquiring various skills that are, you know, step-specific

**[00:15:58 – 00:16:02]** through books and publications or design examples, specifications,

**[00:16:02 – 00:16:05]** best practices, et cetera.

**[00:16:07 – 00:16:10]** But in this new world, you know, in this new agentic

**[00:16:10 – 00:16:14]** world, we have to really start thinking differently, right?

**[00:16:14 – 00:16:19]** We have to move away from traditional workflows, traditional

**[00:16:19 – 00:16:25]** scripting, and actually introduce a more outcome-based mindset, right?

**[00:16:25 – 00:16:30]** And that is the goal that we are jointly pursuing with Synopsys.

**[00:16:30 – 00:16:34]** We announced our partnership with Synopsys last year, and we have

**[00:16:34 – 00:16:39]** been working since then to try and come up with a unique solution that

**[00:16:39 – 00:16:44]** combines the Synopsys Agentic Stack with the MaxR discovery platform.

**[00:16:44 – 00:16:48]** So what this would mean is the Synopsys agents and tools and

**[00:16:48 – 00:16:53]** their workflows would be available through a discovery platform.

**[00:16:54 – 00:16:57]** And we're really glad to have Synopsys as a

**[00:16:57 – 00:16:58]** partner on this journey.

**[00:16:58 – 00:17:02]** We are already seeing some real transformative outcomes

**[00:17:02 – 00:17:04]** with the solutions that we have put together so far.

**[00:17:04 – 00:17:09]** And I'm going to show you a couple of those in the following slides.

**[00:17:10 – 00:17:12]** But before we get to those demos, I wanna just show you

**[00:17:12 – 00:17:15]** behind the scenes on a real life use case, right?

**[00:17:15 – 00:17:18]** So very simplistic use case, but just to show you what happens

**[00:17:18 – 00:17:21]** when these agents are working through and when even Discovery

**[00:17:21 – 00:17:25]** Engine is looking at a prompt, what does it do in the background to try

**[00:17:25 – 00:17:26]** and run the whole thing through?

**[00:17:26 – 00:17:29]** So this is a simple prompt.

**[00:17:29 – 00:17:32]** The intent here is to actually design a microcontroller for

**[00:17:32 – 00:17:34]** aircraft instrument landing system.

**[00:17:34 – 00:17:37]** On the right hand side, you'll see a very simple prompt here

**[00:17:37 – 00:17:39]** in national language.

**[00:17:39 – 00:17:44]** Once that's provided to Discovery Engine, basically starts to analyze

**[00:17:44 – 00:17:48]** the spec, available IP portfolio, and start to really figure

**[00:17:48 – 00:17:54]** out how do you decompose this into individually executable tasks.

**[00:17:54 – 00:17:58]** Once it does that and starts to assign agents to it, it

**[00:17:58 – 00:18:00]** composes a workflow out of it.

**[00:18:00 – 00:18:03]** So what you see here is now multiple agents that are all

**[00:18:03 – 00:18:05]** informed on their own domain.

**[00:18:05 – 00:18:08]** So, for example, the first agent is going to be writing

**[00:18:08 – 00:18:09]** a technical spec.

**[00:18:09 – 00:18:11]** Because you have the intent, you have to write a spec first.

**[00:18:12 – 00:18:15]** And then that spec can be then used by the next agent

**[00:18:15 – 00:18:17]** to actually write the Verilog code.

**[00:18:17 – 00:18:20]** And then followed by writing the functional validation

**[00:18:20 – 00:18:22]** or PPA analysis, et cetera.

**[00:18:23 – 00:18:26]** This is just an example, but just shows you the thinking

**[00:18:26 – 00:18:31]** and how you start from an intent and what does the engine do to

**[00:18:31 – 00:18:33]** actually decompose this into tasks.

**[00:18:33 – 00:18:36]** How are tasks actually being assigned to specific agents

**[00:18:36 – 00:18:39]** to go through a workflow?

**[00:18:40 – 00:18:44]** Now, one thing I wanted to point out over here is

**[00:18:44 – 00:18:46]** explainability, which is extremely important, obviously,

**[00:18:46 – 00:18:48]** for all the engineering purposes.

**[00:18:48 – 00:18:51]** If you have consumer applications, you're answering questions,

**[00:18:51 – 00:18:51]** OK, that's all great.

**[00:18:51 – 00:18:53]** But when it comes to engineering workflows, you need high

**[00:18:53 – 00:18:57]** predictability, transparency, and explainability in the flow.

**[00:18:57 – 00:18:59]** Because otherwise, if you're stuck, you have certain outcomes.

**[00:18:59 – 00:19:03]** If you're not able to explain how you got here,

**[00:19:03 – 00:19:04]** Meaningless.

**[00:19:04 – 00:19:08]** So something, an example over here,

**[00:19:08 – 00:19:13]** this ran through and basically the graph rag technology and the

**[00:19:13 – 00:19:17]** engine built-in basically pointed out, hey, this is the results,

**[00:19:17 – 00:19:20]** this is the power analysis, this is the estimated power consumption,

**[00:19:20 – 00:19:23]** but if you really wanted the real, like more realistic

**[00:19:23 – 00:19:26]** numbers, you should add parasitics and you should add delay.

**[00:19:26 – 00:19:30]** Now this information was actually not provided to the system at all.

**[00:19:30 – 00:19:33]** And for people who understand chip design, this sounds really basic.

**[00:19:33 – 00:19:36]** But the thing is, this system did not know it.

**[00:19:36 – 00:19:40]** But it's all the prior knowledge and the documentation, et cetera,

**[00:19:40 – 00:19:42]** with this multi-hop reasoning is able to actually figure out that,

**[00:19:42 – 00:19:45]** hey, there are ways to improve this further, and it's already making

**[00:19:45 – 00:19:48]** recommendations on how to do that.

**[00:19:48 – 00:19:50]** Now you can imagine this in more complex scenarios, like

**[00:19:50 – 00:19:53]** how it could actually start really thinking through what

**[00:19:53 – 00:19:56]** needs to be done next.

**[00:19:56 – 00:19:59]** So with that, I'm going to show you the first demo here.

**[00:19:59 – 00:20:03]** To set this up real quick, this is a spec to RTL demo.

**[00:20:03 – 00:20:06]** This has actually been done by the Synopsys team in close

**[00:20:06 – 00:20:07]** collaboration with Microsoft.

**[00:20:08 – 00:20:10]** And it basically starts from a specification document and

**[00:20:11 – 00:20:14]** generates an RTL and writes the test benches.

**[00:20:14 – 00:20:15]** I'm going to play that now.

**[00:20:15 – 00:20:19]** The process starts with the design specification file.

**[00:20:19 – 00:20:23]** In the chat prompt, we reference the spec.txt file and ask

**[00:20:23 – 00:20:26]** for RTL to be generated.

**[00:20:26 – 00:20:29]** At this point, the RTL has been created and a unit test

**[00:20:30 – 00:20:31]** process has started.

**[00:20:31 – 00:20:35]** We can see that it generated three test cases for us to run

**[00:20:35 – 00:20:37]** in addition to the sanity tests.

**[00:20:37 – 00:20:41]** Let's look deeper at this last test that failed by using

**[00:20:41 – 00:20:42]** the debug button.

**[00:20:42 – 00:20:47]** Let's rerun the failing unit test with our new edits.

**[00:20:47 – 00:20:50]** Now this test passes.

**[00:20:50 – 00:20:55]** This part of the flow now runs the lint agent, which diagnoses RTL

**[00:20:55 – 00:20:59]** problems with lint check, but also attempts to fix them iteratively.

**[00:20:59 – 00:21:02]** The summary tab opens.

**[00:21:02 – 00:21:08]** It shows us a button to compare the original versus the fixed RTL.

**[00:21:08 – 00:21:13]** Synthesizability is our last quality metric to explore.

**[00:21:14 – 00:21:18]** We can use an exposed agent for a fusion compiler.

**[00:21:18 – 00:21:21]** At the end of processing, we are given a clear summary

**[00:21:21 – 00:21:25]** that we are indeed ready for a synthesis from the synthesis

**[00:21:25 – 00:21:28]** check complete message.

**[00:21:30 – 00:21:33]** Okay, so this is a quick overview, but it just shows you the

**[00:21:33 – 00:21:35]** power of this agent, DKI.

**[00:21:35 – 00:21:38]** They're actually running real synopsis tools through

**[00:21:38 – 00:21:43]** a workflow, producing a measurable outcome here.

**[00:21:43 – 00:21:45]** I wanted to show you another demo here.

**[00:21:45 – 00:21:47]** It's also ready in chip design.

**[00:21:48 – 00:21:50]** And I wanted to pick up here a scenario which is

**[00:21:50 – 00:21:51]** slightly more complex.

**[00:21:51 – 00:21:54]** I want to show you how Discovery Engine actually works, and

**[00:21:54 – 00:22:00]** how the thinking process works, and how is it unlocking decision-making

**[00:22:00 – 00:22:04]** when the humans are pitting a wall.

**[00:22:04 – 00:22:06]** So in this scenario, we all know, for folks who understand chip

**[00:22:06 – 00:22:11]** design, the functional coverage, getting the functional coverage

**[00:22:11 – 00:22:16]** high up, is one of the toughest tasks for any RTL designer.

**[00:22:16 – 00:22:20]** And it makes it even tougher if you did not actually write

**[00:22:20 – 00:22:21]** that original code.

**[00:22:21 – 00:22:24]** You inherited an IP, and you're supposed to now extend the

**[00:22:24 – 00:22:27]** functional coverage in the context of your own subsystem design.

**[00:22:27 – 00:22:31]** So we are picking that as the case here, where it's a legacy

**[00:22:31 – 00:22:34]** IP, very valuable, you want to use it in your subsystem, you want

**[00:22:34 – 00:22:36]** to improve the functional coverage.

**[00:22:36 – 00:22:38]** Humans have already tried it, they are the best, and

**[00:22:38 – 00:22:40]** they hit a bit of a wall.

**[00:22:40 – 00:22:43]** And so that's where Discovery Engine really shines.

**[00:22:43 – 00:22:46]** Like how do you now explore this and how uncover new

**[00:22:46 – 00:22:50]** opportunities, new testing strategies to actually improve

**[00:22:50 – 00:22:51]** the functional coverage, right?

**[00:22:51 – 00:22:54]** So I'm gonna play this video.

**[00:22:56 – 00:23:00]** Discovery Engine continuously analyzes coverage data, failures,

**[00:23:00 – 00:23:02]** and inferred design intent to propose prioritized next

**[00:23:03 – 00:23:06]** actions, test hypotheses, and rationale, helping push

**[00:23:06 – 00:23:09]** through late-stage plateaus without relying on trial and error.

**[00:23:09 – 00:23:11]** On the left, you're seeing Discovery Engine acting as

**[00:23:11 – 00:23:13]** a cognitive layer.

**[00:23:13 – 00:23:15]** It's not replacing execution or automation.

**[00:23:15 – 00:23:18]** It's continuously assessing progress, coordinating

**[00:23:18 – 00:23:20]** across workflows, and asking a key question.

**[00:23:20 – 00:23:23]** Are we still making linear progress, or have we

**[00:23:23 – 00:23:25]** entered a phase where deeper reasoning is required?

**[00:23:25 – 00:23:28]** This is not a static plan, but rather the system re-evaluates

**[00:23:28 – 00:23:31]** as results come back.

**[00:23:31 – 00:23:32]** Here's the first critical moment.

**[00:23:32 – 00:23:35]** The system detects a cognitive bottleneck, ambiguous coverage

**[00:23:35 – 00:23:39]** gaps, conflicting signals, and multiple plausible next steps.

**[00:23:39 – 00:23:42]** Instead of pushing forward with partial confidence, discovery

**[00:23:42 – 00:23:46]** advanced context using graph-based and enters a deep reasoning

**[00:23:46 – 00:23:51]** phase, reconciling testing assumptions in the solution space.

**[00:23:51 – 00:23:54]** This is the kind of sustained analysis that's difficult

**[00:23:54 – 00:23:57]** to maintain manually at scale.

**[00:23:58 – 00:24:00]** Once that bottleneck is resolved, the system doesn't return

**[00:24:00 – 00:24:02]** to a single linear path.

**[00:24:02 – 00:24:05]** It launches multiple hypotheses, explorations in parallel,

**[00:24:05 – 00:24:09]** feeding results back into a shared context, replanning and adapting.

**[00:24:09 – 00:24:12]** Explore, reason, propagate, replan.

**[00:24:12 – 00:24:15]** This loop is what allows progress to continue where traditional

**[00:24:15 – 00:24:17]** flows often slow.

**[00:24:17 – 00:24:20]** To ground this, let's review two key metrics, coverage

**[00:24:20 – 00:24:23]** improvement and the breadth and depth of test bench changes.

**[00:24:23 – 00:24:26]** These metrics aren't meant to be absolute benchmarks.

**[00:24:26 – 00:24:28]** They illustrate how the engine accelerates progress

**[00:24:28 – 00:24:31]** on a reused IP, where understanding intent and closing the last

**[00:24:32 – 00:24:35]** gaps typically consumes disproportionate human effort.

**[00:24:35 – 00:24:38]** First is a comparison to initial coverage achieved

**[00:24:38 – 00:24:40]** with reasonable manual effort.

**[00:24:40 – 00:24:42]** In just one attempt, the engine quickly improves coverage,

**[00:24:42 – 00:24:46]** where a human typically struggles with file closure for an IP

**[00:24:46 – 00:24:48]** originating from another source.

**[00:24:48 – 00:24:51]** Next, breadth and depth of test bench modifications are reflected

**[00:24:51 – 00:24:55]** in the number of files and the size and complexity of the changes.

**[00:24:55 – 00:24:59]** Three key takeaways from this demo are, one, it turns indecision

**[00:24:59 – 00:25:01]** into momentum autonomously.

**[00:25:01 – 00:25:04]** Two, it explores the solution space in parallel with

**[00:25:04 – 00:25:06]** reasoning, not retries.

**[00:25:06 – 00:25:10]** And three, it improves as it goes, learning in the loop.

**[00:25:13 – 00:25:17]** Subjit, so what you saw here is like the combination of

**[00:25:17 – 00:25:19]** reasoning and multiple sort of parallel explorations.

**[00:25:19 – 00:25:24]** How is it like really unlocking indecision and

**[00:25:24 – 00:25:26]** or like where the humans are basically hitting a wall?

**[00:25:26 – 00:25:30]** How do you make a step function improvement in the outcomes?

**[00:25:30 – 00:25:34]** Now, like when you're able to actually raise the level

**[00:25:34 – 00:25:36]** of complexity of tasks.

**[00:25:36 – 00:25:38]** That the system can solve.

**[00:25:38 – 00:25:42]** You can start imagining like more complex scenarios, right?

**[00:25:42 – 00:25:44]** Now these are just as a guide, more future looking.

**[00:25:44 – 00:25:47]** These are not possible today, but I don't think it's too

**[00:25:47 – 00:25:51]** far away where we can actually start really thinking about

**[00:25:51 – 00:25:55]** like, if you're doing an SOC, ACO, you wanna swap out an IP.

**[00:25:55 – 00:25:58]** You should be able to just tell what you want to do with

**[00:25:58 – 00:26:02]** a bunch of documentation and a bookshelf that's fully loaded with

**[00:26:02 – 00:26:06]** your prior design data and context information, and it should just be

**[00:26:06 – 00:26:11]** able to autonomously swap out the IP, do the RTL changes, write the

**[00:26:11 – 00:26:16]** test benches, whatever's necessary, be able to produce a clean RTL.

**[00:26:16 – 00:26:18]** Or if you want to retarget my IP block from Foundry A

**[00:26:18 – 00:26:23]** to Foundry B, of course all the Data is already preloaded into the

**[00:26:23 – 00:26:25]** system, so it knows exactly where

**[00:26:25 – 00:26:29]** PDKs are, for example, where the scripts are, what are the

**[00:26:29 – 00:26:31]** criteria, constraints, et cetera, and should be able to autonomously,

**[00:26:32 – 00:26:34]** we will do those tasks, right?

**[00:26:34 – 00:26:39]** So these are all, we believe, it's actually possible to get to this.

**[00:26:39 – 00:26:43]** And if you were to talk about this last year, people were

**[00:26:43 – 00:26:46]** thinking this is at least five to seven years away.

**[00:26:46 – 00:26:49]** Now we believe this is actually much sooner than that.

**[00:26:49 – 00:26:52]** Within the next, you know, 12 to 24 months, I think we

**[00:26:52 – 00:26:57]** should be there where we can start specifying complex tasks like this.

**[00:26:57 – 00:27:01]** And obviously innovation doesn't stop at chip design, right?

**[00:27:01 – 00:27:03]** We are getting into manufacturing as well.

**[00:27:03 – 00:27:07]** There's a lot of opportunities here on like band structure

**[00:27:07 – 00:27:12]** optimization or carrier charge mobilities or thermal management

**[00:27:12 – 00:27:13]** or yield improvement, et cetera.

**[00:27:13 – 00:27:17]** These are all the things that are really, really suitable

**[00:27:17 – 00:27:21]** for generative AI technology, discovery engine type technology

**[00:27:21 – 00:27:25]** to really create big disruption.

**[00:27:26 – 00:27:28]** I'm gonna give you now, moving away from semiconductors a

**[00:27:28 – 00:27:31]** little bit, few more examples in other domains, right?

**[00:27:31 – 00:27:35]** This example is from chemistry material science, a new

**[00:27:35 – 00:27:36]** material discovery.

**[00:27:36 – 00:27:40]** So this is a project that we did in collaboration with

**[00:27:40 – 00:27:43]** Pacific Northwest Laboratory.

**[00:27:43 – 00:27:47]** And this is an example of, the intent here was to actually

**[00:27:47 – 00:27:54]** discover a new battery electrolyte that reduces the use of lithium.

**[00:27:54 – 00:27:56]** Lithium, as you all know, is a rare earth.

**[00:27:56 – 00:27:57]** It has its own supply chain complexities.

**[00:27:57 – 00:28:00]** We want to reduce our dependence on lithium.

**[00:28:00 – 00:28:03]** So the intent was, hey, can we now get a new electrolyte

**[00:28:03 – 00:28:05]** that minimizes the use?

**[00:28:05 – 00:28:07]** Starting with, again, periodic table, of course,

**[00:28:07 – 00:28:08]** like basics, right?

**[00:28:08 – 00:28:12]** 32 million combinations of molecules.

**[00:28:12 – 00:28:16]** And then filtering through this process of HPC simulation

**[00:28:16 – 00:28:20]** and then of course human insight coming in towards the end

**[00:28:20 – 00:28:24]** over here and solving it for stability, availability, et

**[00:28:24 – 00:28:27]** cetera, within nine months.

**[00:28:28 – 00:28:31]** We were able to not only discover the new electrolyte, which

**[00:28:31 – 00:28:36]** was 70% less lithium, but also be able to synthesize in a lab

**[00:28:36 – 00:28:37]** and create a new battery out of it.

**[00:28:38 – 00:28:39]** And it's a functional battery, right?

**[00:28:39 – 00:28:42]** So we are looking at now how do we scale production for this?

**[00:28:42 – 00:28:45]** But it is possible that something like this would have taken

**[00:28:45 – 00:28:46]** years earlier, right?

**[00:28:46 – 00:28:50]** I'm going to give you another example.

**[00:28:50 – 00:28:54]** So this is the one where we wanted to

**[00:28:54 – 00:28:56]** Discover a new immersion coolant for data center

**[00:28:56 – 00:28:59]** applications, right?

**[00:28:59 – 00:29:03]** And these immersion coolants actually use a lot of

**[00:29:03 – 00:29:06]** chemicals that are harmful for the environment.

**[00:29:06 – 00:29:09]** So the intent here was to, can you minimize the use or eliminate

**[00:29:10 – 00:29:13]** the use of PFAS chemicals, which are known as forever chemicals?

**[00:29:13 – 00:29:17]** So here's this prompt that was given to Discovery to

**[00:29:17 – 00:29:21]** identify a novel immersion coolant.

**[00:29:21 – 00:29:24]** And the first thing was doing research, like what is the

**[00:29:24 – 00:29:28]** state-of-the-art, what are the bottlenecks, what is missing,

**[00:29:28 – 00:29:30]** what else is actually available?

**[00:29:30 – 00:29:34]** And based on this research, it actually came up, if you see on

**[00:29:34 – 00:29:37]** the left-hand side of the screen, came up with the actual plan.

**[00:29:37 – 00:29:40]** And this plan of execution came up with the hypothesis,

**[00:29:40 – 00:29:44]** okay, these are the potential

**[00:29:45 – 00:29:50]** molecules that could actually avoid the use of PFAS chemicals.

**[00:29:50 – 00:29:53]** And came up with a plan which actually uses the agents that

**[00:29:53 – 00:29:57]** are also available in the system, and then be able to

**[00:29:57 – 00:29:59]** execute that entire workflow.

**[00:29:59 – 00:30:07]** And produce the molecules that were basically satisfying the criterias.

**[00:30:07 – 00:30:12]** And we sent it to the lab, and we actually got the

**[00:30:12 – 00:30:13]** coolant synthesized.

**[00:30:14 – 00:30:18]** And you can see this actually running a high-performance gaming

**[00:30:19 – 00:30:24]** on top of this, and being extremely effective in cooling the system.

**[00:30:24 – 00:30:27]** It's also another real-world outcome here, right?

**[00:30:27 – 00:30:32]** Switching gears to pharmaceuticals, like healthcare life sciences,

**[00:30:32 – 00:30:35]** this is a problem that our own team, our own office of

**[00:30:35 – 00:30:39]** CTO, basically picked up and they said, how do we improve

**[00:30:40 – 00:30:43]** the efficacy of the monkeypox vaccine, which was the state

**[00:30:43 – 00:30:47]** of the art, this is last year.

**[00:30:47 – 00:30:51]** And so perform the research, figure out what is the latest

**[00:30:51 – 00:30:54]** that's available, where the key bottlenecks are, and it figured

**[00:30:54 – 00:30:59]** out that there is a protein binding efficiency issues, the reason

**[00:30:59 – 00:31:00]** why the efficiencies are low.

**[00:31:00 – 00:31:03]** So if you figure out new proteins, you can actually

**[00:31:04 – 00:31:05]** improve the efficiency.

**[00:31:05 – 00:31:07]** So that was its own analysis.

**[00:31:07 – 00:31:10]** And based on that, it actually came up with a workflow.

**[00:31:10 – 00:31:12]** If you see on the right hand side, there's like

**[00:31:12 – 00:31:13]** 10 agents and 12 tools.

**[00:31:13 – 00:31:18]** That it picked up to do this, to implement this workflow.

**[00:31:18 – 00:31:21]** There are some senior, more qualified agents, if you will,

**[00:31:21 – 00:31:23]** postdoc level, the virtual scientists, human inputs.

**[00:31:23 – 00:31:26]** As you can see, there's hierarchy being created.

**[00:31:26 – 00:31:30]** And they basically ran a full run, which ran for

**[00:31:30 – 00:31:32]** about five hours or so.

**[00:31:32 – 00:31:36]** And it actually resulted in eight different proteins that could

**[00:31:36 – 00:31:39]** actually solve the problem, right?

**[00:31:39 – 00:31:41]** Actually improve the binding efficiency, thereby improving

**[00:31:41 – 00:31:43]** efficiency of the vaccine.

**[00:31:43 – 00:31:45]** And what you see in the right-hand side is a time

**[00:31:45 – 00:31:49]** lapse, obviously, and how the workflow actually got executed.

**[00:31:49 – 00:31:51]** So this is all real, you know?

**[00:31:51 – 00:31:54]** These are not research papers anymore.

**[00:31:54 – 00:31:56]** I wanna point out a couple of examples

**[00:31:56 – 00:31:57]** on the physics side as well.

**[00:31:57 – 00:32:01]** So this is with our partner, PhysicsX.

**[00:32:02 – 00:32:05]** Basically, the problem here was to optimize or create

**[00:32:05 – 00:32:09]** a novel type of artificial heart valve which reduces

**[00:32:09 – 00:32:14]** the complications due to the blood damage and the blood actually, due

**[00:32:14 – 00:32:17]** to the mechanical pumping action, the blood was actually getting

**[00:32:17 – 00:32:18]** damaged in the traditional design.

**[00:32:18 – 00:32:21]** So, how do you now come up with a novel design so we

**[00:32:21 – 00:32:23]** can reduce the damage?

**[00:32:23 – 00:32:26]** And so they took that problem, they have a unique solution

**[00:32:26 – 00:32:29]** of actually creating AI surrogate models of the object you are

**[00:32:29 – 00:32:33]** looking to design so the inference is significantly faster.

**[00:32:33 – 00:32:36]** And so in this case, you would see, like this design is actually

**[00:32:36 – 00:32:40]** dynamically changing because it's just an inference call.

**[00:32:40 – 00:32:43]** And over 50 iterations,

**[00:32:43 – 00:32:47]** On this design, we're done under a minute to actually

**[00:32:47 – 00:32:50]** come up with the appropriate

**[00:32:51 – 00:32:55]** design geometries for this particular wall that actually gives

**[00:32:55 – 00:32:58]** you the enough pressure tolerance.

**[00:32:59 – 00:33:01]** Again, true transformation.

**[00:33:01 – 00:33:03]** Typically for something like this, for every variation,

**[00:33:03 – 00:33:05]** every iteration that you're seeing here, you would actually

**[00:33:05 – 00:33:09]** have to run simulators, generate data, visualize it and see,

**[00:33:09 – 00:33:10]** oh, I'm not there.

**[00:33:10 – 00:33:13]** So then you go back and change parameters, do it all over again.

**[00:33:13 – 00:33:16]** This thing is all happening real time.

**[00:33:17 – 00:33:20]** Gonna show you another example.

**[00:33:21 – 00:33:24]** This is an example from our own Microsoft Surface team.

**[00:33:24 – 00:33:27]** As you know, all the consumer devices, compute devices,

**[00:33:27 – 00:33:28]** they all have cooling systems.

**[00:33:28 – 00:33:31]** Thermal management is one of the toughest problems to

**[00:33:31 – 00:33:33]** solve because of, it's an obvious reason, you have heat

**[00:33:34 – 00:33:36]** dissipation, you've got noise in the system, you've got form

**[00:33:36 – 00:33:40]** factor, you've got cost, all of those factors have to be optimized.

**[00:33:41 – 00:33:45]** So they took initiative, along with PhysicsX as our partner, to design

**[00:33:45 – 00:33:48]** a new fan, cooling fan, right?

**[00:33:49 – 00:33:51]** So here, in the scenario that you're gonna see in the video

**[00:33:51 – 00:33:55]** I'm gonna show next, is that they had a legacy design,

**[00:33:56 – 00:33:59]** and the feedback that came from the customers for the initial

**[00:33:59 – 00:34:01]** user group is that it is too noisy.

**[00:34:02 – 00:34:05]** And they wanted to now optimize the design.

**[00:34:06 – 00:34:08]** With that feedback, okay, so I'm going to just play the video.

**[00:34:08 – 00:34:09]** I'm going to walk through this.

**[00:34:09 – 00:34:11]** So this is an interaction that's happening between UX

**[00:34:11 – 00:34:13]** teams and the design team.

**[00:34:13 – 00:34:16]** The UX team basically provided this feedback saying, hey,

**[00:34:16 – 00:34:18]** this is the new spec.

**[00:34:18 – 00:34:22]** The noise is too much, and we need to have a redesign on the fan.

**[00:34:22 – 00:34:26]** So this engineer basically

**[00:34:26 – 00:34:30]** Just normal chat, they're gonna take this feedback, they're

**[00:34:30 – 00:34:34]** gonna take the original spec, they're gonna take the feedback,

**[00:34:34 – 00:34:36]** and they will come to discovery.

**[00:34:36 – 00:34:40]** They will see that next, that this all was then ported into

**[00:34:40 – 00:34:42]** Microsoft Discovery.

**[00:34:42 – 00:34:46]** And a prompt is a natural language prompt, basically saying,

**[00:34:46 – 00:34:48]** hey, just look at all these specs.

**[00:34:48 – 00:34:53]** And give me and generate like, you know, n number of different cooling

**[00:34:53 – 00:34:58]** fan designs that can satisfy these old and the new spec, okay?

**[00:34:58 – 00:35:01]** Now, this is running Physics X agents in the background,

**[00:35:01 – 00:35:06]** the CFD agents, and actually trying to give you optimized designs.

**[00:35:06 – 00:35:09]** And in this case, he's going to just click on the optimization

**[00:35:09 – 00:35:12]** results, and you can now visualize some of these designs.

**[00:35:12 – 00:35:15]** Each of the triangles that you see on the map here is

**[00:35:15 – 00:35:19]** trying to optimize between the max flow rate and the noise.

**[00:35:19 – 00:35:21]** It's actually a physical design.

**[00:35:21 – 00:35:27]** So you can see the Pareto Frontier, and the user can now pick the ones

**[00:35:27 – 00:35:30]** that they found most interesting.

**[00:35:30 – 00:35:34]** And they can actually look at how they can visualize it.

**[00:35:34 – 00:35:38]** And you can see, like, four designs have been analyzed

**[00:35:38 – 00:35:40]** simultaneously, right?

**[00:35:40 – 00:35:43]** You would be happy with the designs that you see over

**[00:35:43 – 00:35:44]** here, or you may not be.

**[00:35:44 – 00:35:47]** Then you go back, and then you iterate some more, and

**[00:35:47 – 00:35:49]** you can ask the co-pilot to actually change parameters

**[00:35:49 – 00:35:52]** in whichever direction you want.

**[00:35:52 – 00:35:54]** And this is the reality. But the thing is, this is all

**[00:35:54 – 00:35:55]** happening real time.

**[00:35:55 – 00:35:59]** Starting from the high-level specification to get to something

**[00:35:59 – 00:36:01]** like this would otherwise be weeks of work.

**[00:36:02 – 00:36:04]** So now that that's done, it's going to basically say, okay,

**[00:36:04 – 00:36:07]** you know what, I picked this design, it looks good.

**[00:36:07 – 00:36:08]** These four are my finalists.

**[00:36:08 – 00:36:11]** I'm going to now generate detailed report for these fans.

**[00:36:11 – 00:36:14]** And it comes back to Discovery and say, hey, give me a detailed

**[00:36:14 – 00:36:18]** report for all of these, run the real-life simulation for these.

**[00:36:18 – 00:36:21]** So now you saw the thousands of different triangles over

**[00:36:21 – 00:36:24]** there, pick the ones that are most interesting to you.

**[00:36:24 – 00:36:27]** Visualize, and then now you say, OK, now generate

**[00:36:27 – 00:36:28]** a detailed report for me.

**[00:36:28 – 00:36:31]** And this is where real simulation is being run in the background.

**[00:36:31 – 00:36:35]** Of course, time lapse here, and then produces the final report.

**[00:36:35 – 00:36:39]** If all is good, your fan design is done.

**[00:36:40 – 00:36:47]** So this is real, true disruption for engineering design.

**[00:36:48 – 00:36:50]** And this is my last example.

**[00:36:50 – 00:36:53]** I saved the best one for the end.

**[00:36:53 – 00:36:57]** Same Surface team, but basically said, okay, I've designed the fan.

**[00:36:58 – 00:37:02]** Looks good, but fans are expensive, and they have their own form

**[00:37:02 – 00:37:02]** factor limitations.

**[00:37:02 – 00:37:05]** It adds cost to the system, et cetera.

**[00:37:05 – 00:37:09]** What if I were to take my traditional heat sink design,

**[00:37:09 – 00:37:13]** A passive heat sink and enhance it to satisfy my my criterias

**[00:37:13 – 00:37:15]** right heat dissipation criterias.

**[00:37:15 – 00:37:18]** So on the left you see is the original design that was

**[00:37:18 – 00:37:22]** done by human right and then they give the same problem

**[00:37:22 – 00:37:27]** to generative AI and this is the design it produced right.

**[00:37:27 – 00:37:32]** So this looks very organic, very nature-inspired design.

**[00:37:32 – 00:37:37]** This is something that is really beyond human intuition, right?

**[00:37:37 – 00:37:39]** But the kicker is this.

**[00:37:39 – 00:37:42]** We actually did additive manufacturing and have a

**[00:37:42 – 00:37:44]** part that's actually working.

**[00:37:44 – 00:37:46]** It's not just functional.

**[00:37:46 – 00:37:49]** It's actually 25% better in heat dissipation than the design

**[00:37:49 – 00:37:53]** on the left, and it's lighter.

**[00:37:53 – 00:37:54]** So this is the truth.

**[00:37:54 – 00:37:56]** Of course, this has manufacturing challenges, as you can see,

**[00:37:56 – 00:37:59]** but those can be solved and, of course, get guided better.

**[00:37:59 – 00:38:02]** But the thing is, if you really leave this technology

**[00:38:03 – 00:38:06]** unconstrained, it can really produce results

**[00:38:06 – 00:38:10]** and outcomes that humans can't even comprehend, right?

**[00:38:10 – 00:38:12]** This is the true power of AI.

**[00:38:12 – 00:38:16]** So this agent tech revolution is on us, and we want to just

**[00:38:16 – 00:38:19]** see where this platform takes us, and this is just the beginning.

**[00:38:20 – 00:38:22]** Thank you very much.

**[00:38:25 – 00:38:27]** Thank you, Prashant, for the excellent talk.

**[00:38:27 – 00:38:31]** We now have time maybe for just one question.

**[00:38:33 – 00:38:36]** So for the last image you are showing, so you are feeding

**[00:38:36 – 00:38:42]** the same parameters and data that you had on the human-made design?

**[00:38:42 – 00:38:46]** So this was specifically a topology optimization exercise

**[00:38:46 – 00:38:49]** that, you know, that the team basically did with the same

**[00:38:49 – 00:38:53]** parameters, but it left it unconstrained for many others,

**[00:38:53 – 00:38:55]** like for physical, obviously, you know, the parameters,

**[00:38:55 – 00:38:56]** it left it unconstrained.

**[00:38:56 – 00:39:00]** But from a cost and thermal dissipation perspective, those

**[00:39:00 – 00:39:03]** constraints were the same for both.

**[00:39:03 – 00:39:06]** Multimodal deep learning models.

**[00:39:06 – 00:39:08]** You had to write models from scratch that would

**[00:39:08 – 00:39:10]** do the optimization.

**[00:39:10 – 00:39:12]** What was the modeling approach behind it?

**[00:39:12 – 00:39:14]** So I don't have all the detailed information on this.

**[00:39:14 – 00:39:16]** This was actually done by our Surface team, but I can

**[00:39:16 – 00:39:19]** certainly provide you the information on this.

**[00:39:19 – 00:39:21]** This was also done sometime last year.

**[00:39:21 – 00:39:26]** But yeah, I can certainly provide you the information on that.

**[00:39:26 – 00:39:28]** Yeah, thanks again, and that's all for today.

**[00:39:28 – 00:39:30]** Thanks again, Prashant.

**[00:39:30 – 00:39:31]** Thank you.

**[00:39:31 – 00:39:33]** Yeah, let's get another round of applause.