TECH

How artificial intelligence exceeded Samsung’s expectations

For Samsung, designing new processors for its upcoming smartphone just got a lot faster. The South Korean manufacturer's System LSI division decided to take a hands-on approach and integrate Anthropic’s Claude Code platform into its semiconductor design and verification processes.

The result of this technological bet is impressive, to say the least, and demonstrates that these new tools have changed the game. Extremely complex tasks that used to drag on for over a month are now being completed in a matter of days, proving that automation is the way forward in hardware development.

Despite the breakneck speed and incredible progress, oversight by human engineers remains a vital piece of the puzzle. The machine still makes some serious technical blunders that, without human intervention and supervision, could derail the manufacturing of thousands of devices.

Designing an integrated system is usually a painful, lengthy, and highly meticulous process. However, in a recent project focused on a chip featuring 64 extremely complex data paths, the platform managed to create a virtual test environment and perform all necessary validations in just two days. Under normal working conditions, the same team would have taken over a month to complete this heavy development phase.

What is most intriguing about all this is that the artificial intelligence performed incredibly well even when crucial components were missing. Faced with delays in the delivery of documentation and the codebase for a memory controller, it used available specifications to insert temporary virtual blocks and analyze the structure, saving the technical team a significant amount of time.

To get a clear picture of this technology's impact within the tech giant's labs, here are some of the major milestones recently achieved:

-Completion of complex hardware verifications about 15 times faster than traditional engineering methods. Junior engineers can now create virtual emulator models for external devices in a single day—a process that previously required weeks of study.

-Autonomous implementation of structural replacement blocks to test incomplete or delayed data circuits.

-A drastic and immediate reduction in the learning curve for new developers joining the company.

Samsung’s urgency to implement these advanced technologies in its daily operations is no coincidence. Currently, the division responsible for developing Exynos processors has around 6,000 employees—a figure that seems impressive until you look at its biggest rival. Qualcomm employs nearly 52,000 people, creating a massive gap in capacity for developing new components.

After seeing profits slip and launching heavyweights like the recent Galaxy Z Fold 8 exclusively with Snapdragon processors, the manufacturer had to step up its game. Adopting this new tool, alongside well-known platforms like Google Gemini and ChatGPT, is part of a transformation plan designed to finally streamline production.

Regarding Samsung's adoption of AI in chip design...In the two cases cited in the report, the company reportedly found that AI "significantly" reduced development time and assisted the engineering team.

One task involved verification services for a custom System-on-a-Chip (SoC)—such as reviewing internal data connections based on simulation stages and test environments—a process that would normally take over a month but was completed in just two days.

In the second example, an employee used AI to develop USB device models for a scenario emulator. This activity typically requires weeks of manual work but was finished in a single day.

On the other hand, the report also notes that Claude Code made some procedural errors that put the team on alert. Cited instances include "hiding" the resolution of errors the AI ​​was supposed to fix, reverting tasks that had already been completed, and attempting to modify code that was meant to remain intact.

For this reason, the team views the tool as an assistive aid rather than a replacement or a system capable of autonomous operation. The report further states that engineers review all platform outputs to prevent errors or "hallucinations"—a practice that is particularly critical in an environment with little margin for error, such as the chip industry.

The human touch remains essential in these processes...In the specialized world of hardware, things work very differently than in conventional software applications. If a processor enters mass production with a physical defect, no over-the-air system update can save it. That is precisely why professionals treat generative platforms strictly as assistants, required to meticulously review every line of generated output.

Some of the errors made autonomously serve as prime examples that the technology does not yet fully grasp the context of modern hardware. In one particularly absurd instance, instead of fixing the root cause of a structural problem, the system merely masked the failure by changing the label from "error" to a simple "information" notification. On other test occasions, it even attempted to modify critical circuit designs it wasn't authorized to access, forcing the humans to keep a very tight rein on it at all times.

mundophone