CXMT planned to use stolen Samsung IP to develop its DRAM, court hears — former Samsung engineer who jumped to Chinese memory maker now behind bars

ChangXin Memory Technologies (CXMT), China’s largest maker of dynamic random access memory (DRAM), planned to obtain Samsung’s IP related to its memory process technologies in a bid to develop its own production nodes, according to testimony given in a South Korean criminal case involving leaked Samsung technology, reports Maeil Business. The company’s managers allegedly had no intention to develop an in-house process technology from scratch, so obtaining IP from Samsung was a strategic decision.

The intention was a key part of CXMT’s business plan rather than an opportunistic attempt to acquire Samsung’s technology after the start of its own DRAM process development, according to the testimony of a former Samsung engineer identified as Jeon, who worked at Samsung for 28 years before jumping ship to CXMT sometimes around 2016. This April, Jeon was sentenced to seven years in prison for illegally obtaining a 600-step Process Recipe Plan (PRP) information concerning Samsung’s DRAM fabrication processes in general and the 18nm-class node in particular.

“From the beginning, CXMT CEO and CTO tried to develop DRAM by stealing Samsung Electronics’ Process Recipe Plan (PRP) information,” Jeon told the court this week, according to Maeil Business. “CXMT had no research institutes or facilities at the time of its establishment, and had no intention of developing technology through its own research.”

Copying the recipe does not automatically give you a working process. Nonetheless, a detailed process recipe can dramatically shorten development of any production node because it reveals things like the sequence of steps, deposition or etch conditions, anneals, cleans, implant parameters, temperatures, pressures, tool types, and so on. For a memory process, a ready process recipe can remove a huge amount of trial-and-error and reveal how the process integration is supposed to work. Meanwhile, even to ensure that the recipe works as intended, one needs to obtain the same or compatible tools as well as adjust their characteristics accordingly.

However, while a process recipe can potentially help to ramp any fabrication technology, a process recipe itself does not necessarily reveal the complete transistor architecture or physical design. For that, perpetrators also want things like cross-sectional structures, masks/layout information, design rules, device dimensions, capacitor geometry, process window, and so on. Such details are not easy to get for a brand-new fabrication technology, but they can be obtained by reverse engineering of shipping products. By combining process recipe with reverse engineering, experienced DRAM engineers can reconstruct much of the technology. Meanwhile, process integration engineers can save years of trial-and-error with actual fab tools.

Normally, it takes DRAM companies around 3 – 5 years to design a process node from device development to yield ramp. For a new player, development time from scratch will probably be around five years rather than three unless pre-competitive IP is licensed or certain heavy-lifting is done elsewhere. In fact, the Maeil Business report claims it took Samsung five years to develop its 18nm-class node.

Meanwhile, CXMT began mass production of DDR4 SDRAM in 2019 using a 22nm-class node. Mass production of memory using a 18nm-class process technology began in 2023. Such a timeline raises a question of whether CXMT’s 22nm-class G1 was derived/adapted from the knowledge of Samsung’s process technologies and/or their process recipes, the more advanced 18nm-class node, or did the allegedly obtained Samsung IP primarily help CXMT develop the later 18nm generation? In any case, there is a court ruling that CXMT stole IP from Samsung to boost its own memory production capability.