In the fast-paced world of AI chip design, time is a decisive factor in staying competitive. For Tenstorrent, a leader in AI hardware development, accelerating the verification process while ensuring that no bugs are left uncovered before tapeout was a key consideration. Tenstorrent partnered with ChipStack to accelerate their verification process to drastically reduce development time for their AI chips.

The results not only demonstrated substantial time savings to achieve similar coverage, but also enabled finding of bugs that human experts may have missed. Notably, the LLMs that power ChipStack’s software were run completely on Tenstorrent’s own TT-LoudBox machines, demonstrating preparedness of their hardware for production workloads.

The Challenge: Lengthy Manual Verification

As AI chips become more complex, the verification phase becomes an increasingly time-consuming and critical bottleneck. In fact, over 60% of all the engineering hours spent on developing a chip go towards design verification. This was similar at Tenstorrent as well. For example, verifying one of their complex blocks (Design 1, hereafter) using formal verification required around 104 hours of expert engineering work. Another complex block required instantiating 10s of complex verification IPs (VIPs) apart from human-written tests (Design 2, hereafter), taking over 40 hours of manual work by an expert formal engineer.

The Solution: ChipStack’s AI-Powered Verification Tools.

ChipStack’s suite of verification tools consist of several AI agents, including the Formal Agent, which Tenstorrent decided to deploy for the above use case. The Formal agent can take the following steps to assist in verification, all while taking feedback and assistance from the end user:
● Design Intent Extraction from RTL/Specifications, with the ability to export as specification files
● Formal Testplan Generation, automatically creating assertion, assumption, and cover scenarios
● Formal Testbench Generation with VIPs (Verification IPs), and SVA implementation of the above testplan
● Testbench Run and Debug analysis, integrated directly within ChipStack’s VSCode extension

The Results: Faster Time to Verification, Unique Bugs Found

Throughout the evaluation period of three months, ChipStack enabled Tenstorrent’s engineering team to verify three critical design blocks more efficiently than before. Over 100 hours of manual verification work were saved, while ChipStack's tool identified multiple bugs, including one that had gone undetected through manual testing. This marked a significant milestone in Tenstorrent's verification workflow, showcasing how automation can amplify engineering productivity.
ChipStack’s solution allowed engineers to reduce verification time by up to 4x on certain designs, delivering the same or better coverage without compromising the bugs found. In one of the designs, an engineer spent less than an hour on tasks that would have previously taken an estimated 30 hours manually.

1. Design 1: Tenstorrent's manual verification process for this highly complex block took around 104 hours. With ChipStack, the time was reduced to under 20 hours, achieving similar or better coverage. Both manual and automated verification uncovered the same set of bugs, but ChipStack’s speed allowed Tenstorrent to address them much sooner in the development process.

2. Design 2: This complex block that connected two important protocol interfaces required 10+ VIPs and additional tests by humans. With ChipStack’s tools, which automatically find and instantiate the correct VIPs, humans had to spend no extra time on top of what ChipStack already generated. The produced testbench found several failures that are actively being investigated to see if these are legitimate design bugs.

3. Design 3: In this case, ChipStack's AI-powered tooling discovered an issue still being discussed by Tenstorrent’s engineering team—an issue that an experienced engineer had missed during manual verification. This level of insight showcased the power of AI in catching edge cases that might otherwise go unnoticed.

   
   

Varun Ramesh, Principal Engineer for Formal Verification at Tenstorrent, says:

"We’ve successfully reproduced results across several design blocks and are actively working to scale this across more projects. The tool has proven invaluable in enabling engineers to understand complex designs, create high-quality formal testplans and testbenches. Going forward, using ChipStack can make formal verification more accessible to designers and DV engineers alike.”

Ease of Deployment and Data Security

One of the critical challenges with AI-driven solutions, especially in industries like semiconductor design, is ensuring secure deployment that safeguards sensitive data. For Tenstorrent, maintaining extensive on-premise infrastructure for chip development meant that any third-party tool integration needed to be done with the highest security and compatibility standards.

ChipStack’s solution was designed with these requirements in mind, enabling deployment within Tenstorrent’s secure environment without compromising data integrity or introducing external risks. This was achieved without any data leakage or need to share information externally, aligning with Tenstorrent’s data privacy policies.

"The deployment of ChipStack’s solution was seamless. The on-premise setup integrated smoothly with our existing infrastructure, which allowed us to maintain complete control over our data while leveraging ChipStack’s advanced verification tools,"

says Sam Huffman, Senior Principal Engineer at Tenstorrent.

By deploying directly onto Tenstorrent’s infrastructure, ChipStack ensured that the entire verification process—along with the data and results—remained within the secure boundaries of the company’s network. This approach provided Tenstorrent with confidence in using a generative AI-based tool without the risk of exposing sensitive design information.

Using AI to build AI Chips: Running ChipStack on Tenstorrent’s Hardware

This engagement went one more step further to cement the collaboration between the two companies. ChipStack’s software that uses Large Language Models (LLMs), which are typically run on GPUs from companies like Nvidia. In this case, however, the underlying LLMs were successfully run on Tenstorrent’s own TT-LoudBox hardware.

Equipped with four Tenstorrent Wormhole™ n300s cards for a total of eight Wormhole™ Tensix Processors, the TT-LoudBox machine uses an Ethernet-based mesh topology and can expand to achieve a 96GB memory pool. This enabled ChipStack to run an LLM with 70 billion parameters on the machine while serving multiple simultaneous users. Additionally, Tenstorrent’s own software stack was used to power the inference serving on a larger cluster of LoudBoxes managed by Kubernetes, further showcasing Tenstorrent’s readiness for production use cases.

Conclusion

Tenstorrent’s partnership with ChipStack highlights the power of AI-driven verification and a glimpse into the future of how DV will be performed. By automating critical tasks and reducing the time to verification, ChipStack enabled Tenstorrent to focus on its core innovation while maintaining high-quality standards.

"ChipStack’s tools bring a much needed productivity boost to design verification, a core bottleneck in our workflows. Internally, we’ve seen significant interest from engineers in using ChipStack in their day-to-day workflow, hence we have decided to deploy ChipStack in production and have started to explore other capabilities of their platform."

says Divyang Agarwal, VP of RISC-V Cores at Tenstorrent.

For AI chip companies looking to accelerate their development cycles, ChipStack offers a proven, powerful solution.