AI Design In Korea

Applied research in AI taking place under the auspices of institutions like Samsung AIT and KAIST.


Like many in the semiconductor design businesses, Arteris IP is actively working with the Korean chip companies. This shouldn’t be a surprise. If a company is building an SoC of any reasonable size, it needs network-on-chip (NoC) interconnect for optimal QoS (bandwidth and latency regulation and system-level arbitration) and low routing congestion, even in application-centric designs such as 5G basebands. But what about artificial intelligence (AI) chips? When it comes to AI, US and Chinese companies seem to dominate the conversation. Occasionally, there is some buzz about Europe, but Korea never seems to come up. Are they just not interested?

Not at all. They’re very active, but most of what they are doing is application-centric or applied research. Through my literature searches I haven’t found as many efforts into core research like  you see from US and China universities and “superscaler” companies like Google and Baidu. But my company is working with Korean engineering groups creating chips built around AI accelerators and they tend to be very big and have highly complex dataflow requirements. These kinds of devices absolutely need to use a NoC.

Based on published papers, the majority of AI development is in Samsung AIT (Advanced Institute of Technology) and in KAIST, the Korea Advanced Institute of Science and Technology. Samsung uses its AI technology extensively in its own products, such as smart speakers and voice-controlled remotes. Additionally, they offer a range of ADAS products, including NPUs embedded in their Exynos processors. Still, it’s not easy for the average person to see how AI is evolving in Korea because so much of it doesn’t appear in worldwide AI academic or industry forums.

However, a little digging around does turn up more information. Samsung AIT has a paper on content-aware eye-tracking for 3D displays that do not depend on specialized glasses. The method tracks eye-pupil centers using a visual camera and near-infrared LEDs. From this, it recognizes and tracks facial features and can adapt to different lighting conditions, and if the subject is wearing eyeglasses.

They also have work on augmented reality for wearable displays, automotive head-up displays, light-field displays, and hologram displays. In wearables, they have a paper on detecting blood glucose through non-invasive skin-sensing.

On the device front, KAIST has developed a low power GAN (Generative Adversarial Network) chip, mainly targeting mobile applications. The chip can train the GAN neural net on the mobile device without needing to go to the cloud and is nearly 5X more energy efficient than previous deep neural nets. The institute has demonstrated GAN applications on mobile devices, such as face-to-Emoji conversion, super-resolution imaging, and deep fake detection, which should further enhance user appeal.

The KAIST link provides details about chip architecture. A more comprehensive example is given in a paper from ETRI, the Electronics and Telecommunications Research Institute of Korea. They describe an AI processor they have built that claims it is certifiable to ASIL D, the highest safety level in ISO 26262, which is impressive. The AI structure is based on an array of nano-cores that looks at least superficially like other grid-like structures in different approaches. I’m thinking systolic arrays like what Wave Computing did. Additionally, they claim they also use a number of these cores for fault tolerance through dual modular redundancy, which seems like a clever way to get the maximum value out of these grid structures.

I wish I could tell you more, and I’d certainly be happy to be educated further if you have more information. I do know that we have multiple active customers across Korea using our IP in building their AI solutions. Whatever they’re doing with it, I’m sure we’ll be appearing in many more intelligent products coming out of that country.

You can learn more about Arteris IP applications in AI here.

Leave a Reply

(Note: This name will be displayed publicly)