Samsung Electronics adopted Anthropic's large language model (LLM) "Claude" and has significantly shortened some semiconductor design and verification timelines. About three months after first opening Claude's AI coding tool "Claude Code" to software developers, concrete efficiency gains have appeared on the ground. In customized system-on-chip (SoC—semiconductors that integrate multiple functions such as a processor and memory controller onto a single chip) verification, a task expected to take more than a month was finished in two days, and there was also a case in which a second-year engineer completed development work that could have taken more than a month in one day.
The semiconductor industry says this shift could help offset a shortage of design talent. For example, Qualcomm competes with Samsung Electronics' System LSI division in the mobile application processor (AP) and SoC markets. The System LSI division's headcount is estimated at around 6,000 in the industry. Qualcomm, by contrast, had about 52,000 employees as of September last year, making its overall organization nearly nine times larger.
On Aug. 12, according to ChosunBiz reporting, Samsung Electronics' System LSI division has been using Claude for functional verification of customized SoCs and for early-stage software development for semiconductor projects. After first opening Claude Code to software developers in May, Samsung Electronics expanded its use to specialized semiconductor development tasks. It then broadened the scope further to external Generative AI such as Gemini and ChatGPT, and formalized an "AI transformation," introducing AI across all tasks at all affiliates, from research and development (R&D) to production, marketing, and support.
Claude Code, developed by Anthropic, can read entire program code structures and modify files or execute commands, unlike typical Q&A-style Generative AI. In semiconductor development, it is used to create verification code and build test environments based on circuit design materials, verification programs, and communication specifications. The System LSI division, which designs mobile APs and image sensors, has introduced Claude to cut time spent on repetitive tasks and learning new technologies, boosting the productivity of existing staff.
◇ Verification possible even if design materials are delayed… a task that used to take over a month cut to two days
The System LSI division recently completed, in two days, the setup of a verification environment and the verification itself for a project checking the data interconnection structure of a customized SoC, a process expected to take more than a month. Internally, the work speed was assessed to be about 15 times faster.
The project had stringent conditions because the client required a new semiconductor architecture and it used external intellectual property (IP). Some standardized design materials were missing, and the RTL (Register Transfer Level—design materials that express the operation and data flow of digital circuits in code) for the DRAM controller needed for verification was not released on schedule.
Samsung Electronics filled this gap with Claude-based AI. It fed Claude-based AI with the SoC design information it had secured, the chip's internal communication specifications, and verification IP information from electronic design automation (EDA) vendors. The AI located where verification IP was needed, placed and consolidated it, and created a virtual verification environment and test scenarios. In areas where the DRAM controller design had not yet been released, it consolidated virtual blocks for verification to check core data paths first, and it also identified early errors before the actual RTL was available. The verification target had a complex structure with 64 intertwined data paths, and, according to internal assessments, no manual errors occurred in the process of building the verification environment that could arise from humans performing repetitive tasks.
◇ Development that must use unfamiliar specifications… a second-year hire completed it in a day
There was also a case where AI narrowed the gap in experience and expertise among engineers. To develop software before an actual semiconductor is produced, engineers must virtually implement the operation of Universal Serial Bus (USB) devices such as keyboards and mice in an emulator. What EDA vendors provide is reference code for basic data transfer, so engineers had to learn USB communication specifications and build separate models for each device.
Samsung Electronics assigned this task to a second-year engineer and had the person use Claude Code. The employee had no experience with vibe coding or Claude Code. It typically takes a month to learn the USB communication specifications and develop per-device models.
The employee entered the desired functions and the USB reference code into Claude Code. Claude refined the requirements, suggested implementation approaches and code, and even supported revisions. As a result, building the keyboard and mouse models and confirming their operation was completed in a day, and development of the Android operating system (OS) USB device driver using them was also finished.
◇ SoC losses and reliance on Qualcomm… boosting design workforce productivity with AI
The System LSI division is trying to solve structural business challenges with AI. Park Yong-in, head of the System LSI division (president), said at a management briefing in June that although the division achieved record-high sales in the first quarter this year, it was inevitable that it would post an annual loss due to weakness in the SoC business. In the industry, a smaller design workforce than competitors such as Qualcomm is cited as one of the fundamental reasons for System LSI's poor performance.
The Galaxy Z Fold8 and Fold8 Ultra unveiled last month are fully equipped with Qualcomm Snapdragon, and the Galaxy Watch9 and Watch Ultra2 also adopted Qualcomm chips instead of the previous Exynos. With AI adoption, the System LSI division aims to reduce time spent on repetitive consolidation and verification and on learning standards, focus seasoned engineers on high-difficulty tasks, and enable less-experienced staff to expand their responsibilities more quickly, narrowing such gaps.
This year, Samsung Electronics' DS (device solutions) division has also been expanding the scope of AI use. In March, it said it applied AI to analog and logic semiconductor design, cutting some design times by about 50%. Last month, the memory division disclosed that it had reduced by more than 95% the time to readjust the process design kit (PDK—baseline materials that reflect process information in design) to match process changes and that it is applying this in actual work.
However, there is also an internal view that Generative AI's scope of work and outputs must be tightly controlled. In one verification task, when instructed to fix an error, the AI changed the error message to a general information message instead of addressing the root cause. In another case, when asked to roll back only a specific function, it restored other previously completed tasks as well, and when tasked with analyzing verification results, it even tried to modify RTL, the actual circuit design code. Internally, the assessment was that LLMs did not sufficiently grasp the complex dependencies of hardware description languages.
Software can be updated to fix bugs after release, but it is hard to roll back design defects in semiconductors once mass production begins. Samsung Electronics is gradually raising its level of use by having people define the AI's scope of work and re-verify the results. An industry official said, "LLM-based agents are fast, but if not properly controlled, they can lead to major accidents," and added, "ultimately, they will reduce many of the manual steps people used to perform to shorten overall development time, and engineers will focus on goal-setting and final verification."