Samsung with Claude: “Design and verification shortened to 2 days from a month”

Samsung Electronics’ System LSI division, which designs mobile application processors and image sensors, has been using Anthropic’s Claude to build chip verification environments and emulator models. According to Chosun Biz, on one custom system-on-chip (SoC) the work was planned to take more than a month and finished in two days. Chosun Biz reported this on August 12, and Samsung internally rates the speedup at about 15x.
Samsung opened Claude Code to its software developers in May, then widened access to chip-specific development work, so these results came about three months in. Since then the company has extended use to outside models including Gemini and ChatGPT. It has also made an “AI transformation” official policy, putting AI into every task at every affiliate, from R&D to production, marketing and support.
How the SoC verification ran
Claude Code reads a program’s whole code structure, edits files and runs commands. In chip development, System LSI uses it to write verification code and set up test environments from circuit design data, verification programs and communication specs.
The project was a check of the data connection structure on a customer’s custom SoC, and the conditions were difficult. The customer wanted a new chip architecture and the design used IP from an outside vendor. Some of the standardized design documentation was missing. The RTL for the DRAM controller also missed its scheduled delivery. RTL is the code that describes a digital circuit’s behavior and how data moves through it.
Engineers gave a Claude-based AI the SoC design information they had, the chip’s internal communication spec and the verification IP data from an electronic design automation (EDA) vendor. The model worked out where verification IP blocks were needed, then placed and connected them. It generated a virtual verification environment and test scenarios. Where the DRAM controller design was still missing, it wired in a virtual stand-in block so the core data paths could be tested first. That surfaced early bugs before the real RTL existed. The structure under test had 64 interleaved data paths. Samsung’s internal assessment says the build-out also avoided the manual errors that creep in when a person does that wiring by hand.
For comparison, UST, which runs Claude Code inside its own hardware validation pipeline, said in July that the pipeline cuts validation cycle time by 50% to 70%.
A USB model in a day
The second case involved pre-silicon software work. Before a chip exists, software for it runs on an emulator, and the emulator needs virtual versions of USB devices such as a keyboard and a mouse. EDA vendors only provide reference code for basic data transfer. Engineers have had to learn the USB spec and write a separate model for each device. Samsung says that usually takes a month.
Samsung assigned the task to an engineer in their second year at the company. The engineer had never used Claude Code and had no vibe coding experience. They gave Claude Code the functions they wanted and the USB reference code. Claude then refined the requirements, proposed an implementation, wrote the code and helped revise it. Building and testing the keyboard and mouse models took one day. The engineer then used the models to finish a USB device driver for Android.
6,000 engineers against 52,000
System LSI is hoping AI can help with a staffing gap. Industry estimates put the division at about 6,000 people. Qualcomm, its rival in mobile application processors and SoCs, had about 52,000 employees as of last September, close to nine times as many. In the industry, the smaller design workforce is seen as one of the root causes of System LSI’s weak results.
The results have been weak. At a business briefing in June, System LSI head and president Park Yong-in said the first quarter brought near-record revenue, but the struggling SoC business makes a loss for the year unavoidable. The Galaxy Z Fold8 and Fold8 Ultra, unveiled in July, use Qualcomm Snapdragon chips across the board. The Galaxy Watch9 and Watch Ultra2 also switched from Exynos to Qualcomm. The division plans to use AI to cut time spent on repetitive wiring, verification and learning specs. Senior engineers would then move to harder problems, and junior staff would take on more areas sooner.
Samsung’s semiconductor arm has published other figures this year. In March it said applying AI to analog and logic chip design cut some design time by about 50%. In July the Memory division said it had cut the time needed to recalibrate process design kits by more than 95% and was using the method in production work. A process design kit is the reference data that carries a fab process’s parameters into chip design.
Fix the error, not the message
Samsung’s internal reviews also record failures. In one verification job, the AI was told to fix an error. Instead of fixing the cause, it changed the error into an ordinary info message. In another, the request was to roll back one specific function, and the AI also reverted other work it had already finished. When it was asked only to analyze verification results, it tried to edit the RTL, the actual circuit design code. Samsung’s internal assessment says the model does not fully understand the complex dependencies in hardware description languages.
Software bugs can be patched after release, but a chip’s design flaws are hard to undo once it is in mass production. Samsung has engineers set the scope of each AI task and re-verify the output, and it is expanding use in stages. Samsung treats Claude Code as an assistant, not an autonomous engineer.
“LLM-based agents are fast, but without proper control they can lead to a major accident,” one semiconductor industry official said. “Ultimately this will move toward cutting the many manual steps people used to do, shortening overall development time, with engineers focusing on setting goals and final verification.”
Up next
Grok 4.6 ships in Cursor and Grok Build, Video
Grok 4.6 is live in Cursor and Grok Build today with double rate limits for the first week, and xAI says it now matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. On Frontend Arena it sits 13th with 1553 points.
Related stories
- An Amazon AI project blew through its budget by 860%
- JPMorganChase opens Claude Code to staff beyond engineers
- Goldman limits Claude Code to engineers, cutting off traders
- Meta trims Claude Code use, yet spent over $105M in 28 days
- Barclays wants half its developers on Claude Code this year
- Anthropic’s new AWS gateway puts SSO and spend caps in front of Claude Code
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

