One year after being spun out from Intel, Articul8 AI is pulling out the big guns with its newest benchmarking cred. Specifically, it says its Verilog-capable DSM outperforms by 2x the best open source models and at least matches Google Flash 2.0 and Chat GPT-4o.
All this at a fraction of their computational cost with far fewer parameters needed.
The DSM (Domain-Specific Model) of which Articul8 AI speaks is called A8-Semicon and the company explained its expertise and value in a blog just out. Fortunately, Fierce Electronics got a preview and an exclusive interview with CEO and founder Arun Subramaniyan, former data center and AI group general manager for Intel.
“We’ve had massive scaling in 2024 with a fully integrated software stack for infrastructure providing data and model layers and application API layers,” he told Fierce. “We’re not built on the premise one model would solve problems. We believe multiple models are needed to solve complex problems-- not just general purpose languages but domain-specific models for aerospace or oil or telecom.” Or more, including semiconductor design.
At inception a year ago, Boston Consulting Group had already used Articul8 AI software for multiple clients seeking production-ready platforms.
Then it was customers like Uptycs, Scripps and Invest India. Uptycs, a security provider across laptop and cloud, was pleased with Articul8’s speed and scale for its GenAI engines on its security application. Today, Articul8 has worked with aerospace and energy players. A large telco in Japan and another in Europe want full blown network optimization from Articul8, Subramaniyan said. In all, he said the company has more than two dozen large customers. Intel remains attached as a large customer.
Often, companies will find they need to spend $5 million to $15 million to build a suitable model for their needs, but “most don’t even start,” he said. Articul8 gets them 85% of the way to the finish line, he claimed, for a big cost savings.
Articul8 AI's latest benchmarks against open and closed models
The focus of Articul8’s blog and benchmarking includes a deep dive on semiconductor design involving the use of the Verilog programming language which is deployed for Register-Transfer Level design. Verilog is used to write code that defines how logic gates (there are many billions of gates in modern chips) should behave and interact with each other. There is great value in automating Verilog, but even the best general purpose LLMs “perform very poorly” with a 34% average pass rate, according to the blog.
The blog describes some of how Articul8 trained its expert-level semiconductor DSM, A8-Semicon, including several steps including code debugging. The result was a model that showed twice the performance of open models like the latest DeepSeek-R1, LLaMa-3.3 70 B and Qwen 2.5 Coder 32 B. More specifically, the pass rate for A8-Semicon was twice as high as LLaMA 3.3 70B Instruct.
“By building on top of our DSM-Rs with targeted datasets, businesses can achieve significant improvements in code accuracy and functionality for their own domains,” the blog says. “The same approach can be applied to other specialized domains, such as supply chain management and manufacturing…”
Subramaniyan said the benchmarking showed 2x better accuracy, completeness and correctness with compilation checks. Plus, A8-Semicon used a smaller number of parameters--up to 50 times smaller. Cost savings are achieved because customers don’t need a new data platform. “With existing investments, we can get you value,” he said.
“Everybody claims to do what we do, but we are not an ingredients provider,” he said. “We are a full stack provider.”
Articul8 AI on track for $100M in revenue in a year or 16 months
With eight figures in revenue in 2024, Subramaniyan said Articul8 with its 60 employees is on track to earn $100 million in revenue in a year to 16 months. The company claims it can deploy models for customers in under six weeks at 3-4x lower TCO than traditional AI implementations.
That’s the Articul8 AI essence: Companies don’t need bigger models, they need better models tuned for their data.
Asked why Articul8 is pushing its message with a benchmarking blog now, after DeepSeek has surfaced and more smaller AI players are emerging, Subramaniyan responded this way:
“While we are resolute that evaluating individual model performance in isolation is not the right approach -- and that the industry should focus on system-level performance -- the constant noise around general-purpose model benchmarks has forced us to respond. Even at the model level, domain-specific models (DSMs) prove to be better for real-world applications.
“Models like DeepSeek have amplified the conversation, but they also reinforce the fundamental challenge we recognized early on: general-purpose models alone are not enough for specialized enterprise applications. Our approach to benchmarking is about cutting through the noise to demonstrate what works for complex, real-world challenges like semiconductor design.”