Build3 publishers2 min readPublished
OpenAI and Synopsys build GPT-Synopsys to run chip-design tools in an agent loop
OpenAI and Synopsys are building GPT-Synopsys, a model meant to operate Synopsys chip-design tools, under a multi-year deal announced September 30. No benchmark or launch date is published yet, so design teams can examine the hosted data path but have no result to plan against.
The Engineer · Build desk
What happened
- Engineers would set a goal such as better power, performance and area or a closed timing gap, and agents would run the tools, read results, make changes and repeat until engineers review the outcome.
- OpenAI is licensing Synopsys' EDA software to develop the model.
- Early engagements with semiconductor customers are underway, though the companies have not named them, described the designs being tested, or reported a finished chip-design outcome.
- The two companies will sell the service together and share revenue, without disclosing the split or the pricing.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability Agent steps can be accepted or rejected on the EDA tool's own output, so design teams get to judge the model with the same validation they apply to an engineer's work.
- exposure Piloting the service puts proprietary design data on OpenAI-hosted machines, protected for now only by announcement promises on training and encryption.
- cost Bundling compute, model and tool licenses prices the agent's repeated tool runs inside one contract, so its cost against an engineer's time cannot be judged until a price appears.
- decision With no benchmark against an engineering baseline, a team that wants evidence on its own designs has to produce it inside an early engagement.
The EDA tool itself grades each step the agent takes. RuntimeWire describes a setting where success depends on what the engineering tools report, not just whether an answer sounds plausible [17]. I think the tool-graded loop is the best decision in the plan. Engineers use EDA tools to create, test and verify chips. The hard part, per RuntimeWire, is keeping every automated step inside a process engineers can validate before a design moves toward manufacturing [18]. Because the loop stops for engineer review, the model's changes stay inside that process [4].
The Decoder quotes the announcement's statement of the goal: a model that can "reason about chip design and verification, and to directly operate Synopsys' tools" [13]. Synopsys contributes the tools and its chip-design expertise. OpenAI contributes the models and hosts the service [14]. Synopsys was already building agents for longer tasks through AgentEngineer, its portfolio of AI systems for multi-step engineering workflows [7].
The hosting choice needs the closest reading. The service runs on OpenAI-hosted infrastructure and connects to Synopsys.ai and the Synopsys Autopilot platform [6]. A pilot therefore puts customer design data on hosts OpenAI operates [1]. The companies say that data will not be used to train the model [9]. The Decoder reports it will also be stored encrypted [10]. Both promises come from the announcement. Product terms that a customer could evaluate are not yet public [11].
Compute, the model and the tool licenses come as one bundle [9]. An agent that reruns a tool until it reaches a target spends compute and tool time on each pass. Whether that costs less than an engineer making the same passes depends on how the bundle is priced [8].
"By helping them build better chips, we can build better AI and bring it to more people," Brockman said [3]. According to The Decoder, OpenAI already works with Broadcom on chips built for running AI models [16]. Ghazi says AI could significantly speed up the design process, The Decoder reported [15]. A speed figure would only carry over to another team's chips if it named the workflow, such as timing closure or verification. It would also have to compare the agent with engineers working from the same starting design.
What to watch
- Published product terms that spell out storage, access and enforcement of the no-training promise for customer design data.
- A first benchmark that names the workflow and compares the agent with engineers working from the same starting design.
- A price for the bundled compute, model and tool licenses, alongside a launch date.