Samsung Cuts Month-Long Chip Verification to Two Days With Claude: Report

Samsung Electronics has cut a semiconductor verification task that was expected to take more than a month to just two days using Anthropic’s Claude, according to South Korean media outlet Chosun Biz. 

Its System LSI division used Claude to build a verification environment for a customer-specific SoC involving 64 interconnected data paths. 

The work was complicated by delays in receiving some standard design data, including the RTL (register transfer level) for a DRAM controller, as well as the use of third-party IP.

According to the report, Samsung fed Claude the available SoC design information, internal chip communication specifications, and verification IP from electronic design automation vendors. Claude then identified the required verification IP, connected the components, created a virtual verification environment, and generated test scenarios. 

For the missing DRAM controller, it used virtual verification blocks to examine core data paths and identify errors before the actual RTL became available. 

Samsung found no manual errors during the environment-building process and assessed the overall workflow as roughly 15 times faster.

The effect was also visible in work that would normally require engineers to spend weeks learning unfamiliar specifications. 

The report added that a second-year Samsung engineer used Claude Code to develop and verify virtual USB keyboard and mouse models in one day. The work would have normally taken about a month, including learning USB communication standards and developing separate models. Claude helped interpret the requirements, propose an implementation, and modify the resulting code.

Samsung has been expanding this use of AI beyond Claude Code. The company initially opened Claude Code to software developers last May before extending it to specialised semiconductor development.

It has also brought in external generative AI tools including ChatGPT and Gemini as part of its broader “Great AI Transformation” programme, which covers areas from R&D and production to marketing and support.

However, the report flagged instances where Claude changed an error message instead of fixing the underlying issue, restored unrelated completed work when asked to roll back one function, and attempted to modify RTL when instructed only to analyse verification results. 

Samsung is therefore keeping engineers responsible for defining the scope of AI tasks and re-verifying the output.