AI Won’t Kill Verification IP, But It Will Redefine It

Key Takeaways

• AI will enhance, not replace, verification IP by automating test generation and debug.
• Verification IP’s core value will increasingly lie in trust, accountability, and system-level realism, especially as designs become more complex, multi-die, and security-sensitive.
• AI shifts verification bottlenecks from execution to specification quality, raising expectations for clearer requirements, tighter coverage, and more accurate modeling of real-world behavior.

As chip designs advance to 3nm and 2nm nodes, verification IP is becoming increasingly important for navigating a sea of standards, protocols, and complex integration issues.

Verification on average accounts for 68% of the overall development cycle time. Verification IP (VIP) streamlines routine processes to ensure IP is reliable, ensuring it can be integrated seamlessly with different interface models. But as complex designs evolve from planar SoCs to AI-enabled multi-die assemblies, VIP needs to evolve, as well.

VIP is a pre-verified, reusable simulation model that functions as a testbench component. It is responsible for generating traffic, checking protocols, and monitoring behavior to ensure that a design complies with industry standards such as PCIe, USB, or DDR. By identifying issues early in the process, VIP accelerates the verification cycle. VIP is used heavily with protocols like AMBA, PCIe, USB, MIPI, and memory models.

“Why people purchase VIP from Synopsys or other major vendors comes down to credibility and shared responsibility,” said Varun Agrawal, product manager for protocol solutions at Synopsys. “Buyers are not just acquiring a product. They are trusting the vendor’s reputation and willingness to share in both the risk and accountability associated with the solution. Therefore, when deploying open-source elements like IP, VIP, or code, there are key questions that people consider.”

The main concern with open-source is accountability. If something goes wrong or isn’t as expected, who is responsible for fixing it or providing support? When integrating such technologies into critical projects like chip design, it’s unclear who should receive feedback or be held accountable if issues arise. The addition of agentic AI, in particular, complicates those answers.

“Over the past two decades, many companies have emerged in the industry alongside the three major players — Synopsys, Cadence, and Mentor Graphics (now Siemens EDA). Quite a few smaller companies have appeared as well that often end up relying on the larger vendors,” Agrawal noted. “One key reason is quality assurance. If I purchase from a smaller company, has their product been used successfully by a major company? As a startup planning to launch an internet switch that could be deployed in data centers like Nvidia’s, for example, proper verification is crucial. My first questions for any vendor include: ‘Where have you verified your product? How many customers have taped out using your tools, and are those products functioning well? This is the initial level of validation I require, and I can’t afford to spend time debugging or troubleshooting your product because I have my own deadlines. Can you demonstrate that your product is genuinely reliable?’ Currently, I don’t believe agentic AI can solve this or sustain it in the long term. However, I do see agentic AI making an impact on productivity.”

These considerations underscore the pivotal role of trust and accountability in verification IP, especially as industry standards and vendor reputation influence decision-making for critical projects. As the landscape evolves, it becomes essential to explore how advancements in AI and traditional verification approaches are shaping future practices and expectations.

“Artificial intelligence represents a paradigm in which improvements are driven by accelerating processes, focusing more on learning rather than following programmatic rules,” explained Nandan Nayampally, chief commercial officer at Baya Systems. “Rather than implementing additional rules and procedures, AI is allowed to learn from experience, making data the primary factor influencing its performance. The amount and variety of data — such as words or documents — are crucial to effective learning, and the abundance of linguistic data has contributed to notable progress in language processing. Verification will present unique challenges, particularly regarding how it should be conducted. Should verification rely on an intelligent engine that learns? And if so, what is the source of its learning? Advancements in verification techniques will be necessary and improvement is expected, but ultimately it depends on the quality and quantity of data available for training.”

With these industry shifts and technological advancements, it becomes crucial for the design engineer to understand how verification IP is evolving to tackle practical challenges — especially those relating to security, integration, and the influence of AI-driven tools. As the landscape continues to transform, engineers must focus on how VIP adapts to meet their hands-on needs while maintaining robust reliability and efficiency in day-to-day development work.

Verification IP is becoming increasingly more common to ensure protocols are implemented correctly, and systems are properly connected within a system-on-chip (SoC) with minimal verification effort, noted Jason Oberg, security solutions fellow at Arteris. “At the same time, while extremely beneficial for ensuring functional correctness of integration, VIPs are currently not built to address security-specific weaknesses. Many VIPs are constructed as a set of System Verilog Assertions (SVAs) or monitors that are bound into the design to detect functional integration issues. They are not traditionally, however, designed to detect security issues during integration. This shortcoming is primarily because value-based checking via SVAs and conventional monitors does not adequately cover potential confidentiality or integrity violations to secure design assets where assets can transform value, leak through side channels (power, time, etc.), and end up in locations in the design that may be compromised by an adversary.

Some of today’s approaches use information flow concepts to increase security assurance and help ensure protection of secure assets. For example, Arteris’ Cycuity tools allow security monitors to be generated and bound into design just like VIP, and use unique information flow technology that can more comprehensively detect confidentiality and integrity violations to critical design assets.

VIP’s future
So given the increasing role of agentic AI in verification and IP, will AI make VIP obsolete? Not likely.

“AI is not rendering VIP obsolete or unnecessary,” said Frank Schirrmeister, executive director for strategic programs and system solutions at Synopsys. “Rather, it helps increase productivity. VIP’s connection to IP is similar to how block development relates to block verification. You need to separate the teams that design from the teams that verify and have both work from the same spec.”

But AI and LLMs could make it much easier to develop VIP, said Oberg. “The key to making this work well will be correctly guiding the LLM to create correct and valuable VIP for the VIP consumer. The companies developing and shipping VIP will be best suited to do this, and it would be challenging for VIP consumers to do this on their own today. In the mid-long term, I do think AI/LLMs will get sophisticated enough to generate VIP from more generalized specifications/prompts. This will make it much easier for consumers of IP (and associated VIP) to create VIP on their own without the need for the original VIP vendor. This is all in reference to existing VIP, which is primarily a set of System Verilog Assertions (SVAs). More sophisticated checkers and security monitors are at less risk of becoming obsolete from AI/LLMs due to increased complexity in what they are checking.”

In fact, the idea that AI can replace verification IP misses the point about where the real challenges lie in the verification of hardware designs, according to Nidish Kamath, director of product management, Silicon IP at Rambus. “Verification IP exists to capture decades of protocol nuance, interoperability challenges, and silicon learning. AI tools will empower verification engineers to explore scenarios and spot verification gaps faster and reduce manual effort, but these tools will sit on top of trusted verification foundations.”

William Wang, CEO of ChipAgents, agreed. “AI is evolving quickly, but it is not at a stage where it can replace verification IP. VIP encodes years of protocol knowledge, corner cases, and compliance behavior that production teams still rely on for signoff. What AI can do today is sit on top of VIP, automating test generation, integration, and debug, so engineers spend less time wiring things together and more time closing coverage. In that sense, AI increases the leverage of VIP rather than making it obsolete.”

Changes ahead
Industry practices are adapting as both traditional and AI-driven verification methods continue to evolve, raising questions about how teams will choose and implement verification IP in increasingly complex environments.

From a user’s standpoint, working with verification IP across different organizations can be challenging, especially when it includes both smaller and larger developers. Multiple factors influence the choices. Should the design team use every verification IP available, given that both the IP designer and as many verification experts as possible are involved? How will AI impact that?

The practice of using multiple vendors for verification is common. “Even before AI was involved, everyone wanted to verify with as many different components as possible,” Agrawal said. “Verification is never truly complete. You can never claim to be 100% verified. That’s just not feasible. That’s why people want to explore as many avenues as they can. Is this practical? When you try to manage multiple vendor environments and setups, verifying across all of them becomes challenging. And even then, how do you measure success? Suppose you work with 10 vendors. Can you confidently say your verification is 100% complete? No, so there will always be some uncertainty left. Agentic AI can serve as a second vendor operating in parallel with your primary vendor. It’s similar to hiring an agent alongside your main partner to enhance verification processes. This is where agentic AI adds value. It helps establish correlations to specifications, or to find a point of truth and connect results back to that reference. This could involve generating tests or creating golden models, among other tasks. Ultimately, the technology is poised to significantly augment those aspects.”

As industry perspectives evolve, it becomes clear that the interplay between traditional verification IP and new AI-driven approaches is shaping future practices. This convergence prompts a closer look at how organizations navigate the challenges of verification in complex hardware environments. In the rapidly evolving world of AI development, there will be a number of impacts on verification IP, IP, and verification in general.

Rambus’ Kamath noted that AI already is reshaping verification workflows by accelerating test generation, improving coverage closure, and identifying issues earlier in the design cycle. “What this means for verification IP is higher expectations on coverage and quality thresholds, tighter integration, and a greater emphasis on accuracy and completeness,” he said. “Verification IP will need to model real system behavior more faithfully, especially as AI accelerators push memory, interconnect, and power boundaries.”

In terms of AI’s overall impact on verification, Synopsys’ Schirrmeister sees further changes ahead due to agentic AI. “The most significant change is that bottlenecks are shifting. A recent conference panel featured two vendors and two users, who all agreed that AI will greatly help reduce verification bottlenecks — a point with which I agree. I’m observing that verification stands to improve quality, especially considering the valuable knowledge gained in the process. Think about training a model specifically for PCI Express. You can quickly become an expert in PCIe without constantly relying on human experts for confirmation. This makes the process much more efficient and productive. The main shift is that the challenge is moving from verification toward ensuring clarity in specifications. There’s going to be renewed focus on specification engineering. You’ll need to be much more precise and complete when specifying requirements. Just as using language models requires well-crafted prompts, the accuracy and completeness of specifications will likely become the new bottleneck.

Several sessions at the Design Automation Conference in July will address this topic, discussing how to ensure specs are consistent and how to connect them with front-end processes like MBSE, along with requirements management.

Essential ingredients
As Nvidia CEO Jensen Huang put it, the foundational “flywheels” of underlying tools don’t disappear. Even as new productivity tools emerge, core elements like the use of verification IP remain essential.

“If anything, their importance may increase, becoming an even stronger foundation,” Schirrmeister said. “For instance, logic synthesis plays a key role. When logic synthesis was first introduced, some designers resisted, claiming they could optimize their own table lookups for greater efficiency. But as you advance, those fundamental components remain crucial. The same applies to verification, IP, simulation, emulation, and prototyping; you’ll rely on these tools increasingly, because the problem is NP-complete. Essentially, you’re never truly finished, and it’s difficult to know when your work is actually complete.”

This concept is similar to examining design hierarchy. When advancing to a higher level of abstraction, the foundational elements remain present and essential. “I discussed this topic frequently at recent conferences. The main question people ask is whether these changes are taking something away, but I believe they’re simply transforming things. Many underlying aspects still need to be understood, and we shouldn’t lose valuable institutional knowledge. In broader AI conversations, experienced professionals have developed intuition from years in the field. Their ability to grasp concepts like PCI-Express and assess outcomes from discussions is accurate. In fact, it’s a quick test of understanding that will remain crucial. Intuition, along with suspicion, clarity, and consistency, will only become more important.”

For design engineers, balancing tried-and-true verification methods with the emerging capabilities of AI is a challenge. “Many organizations are eager to explore this area, which has resulted in approximately 50 startups developing agentic solutions for VIP,” said Baya Systems’ Nayampally says VIP. “If the technology were better understood, it would likely be consolidated among a few leading companies or adopted broadly by key silicon vendors. There is clear value in VIP, particularly in potentially increasing efficiency by enabling fewer verification engineers to achieve more rapid progress. However, as verification strategies differ across organizations, a uniform set of solutions has yet to materialize. Significant innovation and continuous learning are taking place, but no single solution or group of solutions has clearly distinguished itself as dominant within the industry.”

Still, AI does not replace verification. “AI should enhance productivity and broaden capabilities,” Nayampally said. “Search remains relevant. As former Google CEO Eric Schmidt noted, the search problem persists because there is rarely a single definitive answer to any query. Artificial intelligence will continue to address tasks that can be improved, automated, or optimized for greater user efficiency. Verification processes remain essential, though methods may evolve. For instance, when querying ChatGPT, responses are generated based on its training data, regardless of speed or accuracy, sometimes omitting critical information. If asked for clarification, the model may revise its response. This highlights the importance of expertise in verification. Understanding how to verify or test results is an ongoing challenge. Verification may increasingly rely on mathematical approaches, yet fundamentally it remains about exploring and managing state space, improving coverage, and ensuring inquiries target the correct areas. These aspects will continue to develop alongside technological advancements.”

Against this backdrop of ongoing innovation and shifting industry approaches, these trends are influencing not only the tools and methodologies used but also the skills and mindsets required of engineers moving forward.

A new approach to training engineers may be needed, even as lower-level engineers move up and learn that layer of skills. “There’s also the need to put people through basic training. If someone relies solely on using an LLM without being able to assess its outputs, it becomes problematic. They’ve never been required to write even a simple VIP themselves. That’s something that needs addressing,” Schirrmeister said.

This means engineers who understand foundational concepts are still of very high value, even if their roles evolve. “I think of it like boot camp. You need to know things like what a transistor or gate is, or how a multiplier works. Otherwise, productivity suffers because you might not evaluate results correctly, even when experts provide additional insight. VIP acts as an expert-approved version of what you’d do with AI, and AI just boosts quality and productivity,” he said.

The main point is that payload engineers are expected to reduce control. “If you examine networking speed, there are both control and payload aspects,” Nayampally said. “Essentially, AI’s role is to lower payload costs, while people manage control costs. Specialized knowledge and expertise will remain crucial, since tools are only as effective as those who use or improve them. Achieving AGI (artificial general intelligence) is still far off, and although we strive toward it, it’s unlikely to happen suddenly.”

While foundational skills and expertise remain crucial in the verification landscape, the evolving roles of engineers and the integration of AI technologies are prompting a shift in how teams approach problem-solving and adapt to new tools.

Elon Musk has publicly stated that the focus now needs to be on asking the right questions, Agrawal noted. “That’s what boot camps and entry-level programs should emphasize — learning to ask the right questions. Over time, mastering this skill becomes increasingly challenging. It’s easier to answer questions than it is to formulate the right ones. This is why it’s becoming even more essential. For example, can you pose the right questions about UVM or about code to arrive at the perfect solution?”

Conclusion
AI is changing the IP industry by shifting value toward deeper domain expertise and system awareness. “As designs become more specialized, IP is increasingly about how components behave together under real workloads,” Rambus’ Kamath said. “AI can help IP vendors design, verify, and optimize faster. On the other hand, AI workloads need high-performant memory subsystems and interconnect, along with data security for the massive amounts of data handled by these systems. The winners in the IP ecosystem will be those who combine AI-driven productivity gains with hard-earned hardware and system insights to accelerate these new computing paradigms.”