Samsung Integrates Claude into Chip Development, Cutting Verification from 1 Month to 2 Days with 15x Efficiency Gain

Samsung Electronics (005930KS) is accelerating the integration of generative AI into semiconductor development. Since opening access to Anthropic’s Claude Code in May, the System LSI division has achieved significant results within just three months. A custom SoC functional verification that previously took over a month was completed in only two days, improving efficiency by approximately 15 times. Additionally, a two-year-experience engineer used AI to finish USB model development—a task that once took a month—in just one day.

In a recent custom SoC verification project, Samsung compressed work that traditionally required more than a month into just two days. Internally, this is assessed as a roughly 15-fold increase in efficiency.

The project adopted a new architecture and external design IP, but some data was incomplete—most critically, the DRAM controller design documents were not delivered on time.

Under traditional workflows, engineers would need to first complete missing documentation, build a verification environment, and then conduct tests item by item, resulting in a lengthy overall process.

After integrating Claude, Samsung fed design information, communication specifications, and EDA verification data into the AI. Claude automatically built the verification environment, generated test scenarios, and used virtual modules to replace missing design documents, enabling early validation of core data paths.

With 64 interwoven data paths involved, the structure was highly complex. Samsung believes AI not only drastically shortened the verification cycle but also reduced human errors caused by repetitive operations.

Junior Engineer Completes One Month of Work in One Day

Beyond shortening chip verification timelines, AI is also reducing semiconductor R&D’s reliance on engineer experience.

Previously, engineers had to develop models for USB devices like keyboards and mice in simulation environments. As EDA vendors provided only basic reference code, engineers needed to study USB specifications and write their own models—a process typically taking about a month.

This time, Samsung assigned the task to an engineer with only two years of experience and instructed them to use Claude Code. After inputting functional requirements and reference code, the AI decomposed the requirements, generated code, and iteratively refined it based on test results.

Ultimately, the development and verification of keyboard and mouse models were completed in just one day, followed by successful development of Android system USB device drivers.

AI as a 'Workforce Amplifier' for System LSI

Samsung’s aggressive AI adoption is also tied to operational pressures faced by its System LSI division. Division head Park Yong-in stated in June that despite record-high Q1 revenue, the full year would likely end in loss due to underperformance in the SoC business. Galaxy Z Fold8, Fold8 Ultra, Galaxy Watch9, and Watch Ultra2 have all shifted to Qualcomm-based solutions.

R&D manpower shortages are also seen as a factor weakening System LSI competitiveness. Facing a workforce gap of nearly 9x compared to competitors, Samsung aims to boost per-engineer productivity using AI—compressing repetitive development, verification, and specification learning time—so senior engineers can focus on complex tasks while helping junior engineers quickly bridge experience gaps.

In fact, Samsung has been steadily expanding AI applications in semiconductor R&D this year. In March, the company revealed design time for certain analog and logic chips was reduced by about 50%. Last month, the memory division reported PDK update and tuning time cut by over 95% with AI assistance.

AI 'Overreach' Remains a Concern

However, as generative AI moves deeper into core chip development processes, risks are emerging. Internal testing at Samsung found instances where AI exhibited 'overreach' or incorrect handling.

For example, when asked to fix an error, the AI directly modified the error message to mask the issue rather than addressing the root cause. When instructed to roll back a specific feature, it sometimes reverted other previously completed work as well.

More notably, during analysis of verification results, the AI attempted to directly modify RTL code for actual circuit designs.

Samsung believes this reflects current large language models’ limited understanding of hardware description languages and complex dependency relationships.

Since chip designs cannot be patched post-volume production like software, Samsung currently sets artificial boundaries for AI operations and conducts human review of all outputs before gradually expanding usage scope.

For Samsung, AI is not intended to replace chip engineers at this stage, but to compress repetitive R&D and verification cycles. AI handles rapid execution, while engineers retain control over goal setting and final validation.