Intel, along with NVIDIA, Broadcom, and Qualcomm, has been investing for a long time to conquer the on-device (embedded) artificial intelligence (AI) market and has released AI chips for edge use, but there is still no clear dominator. This contrasts with the fact that NVIDIA's graphics processing units (GPUs) have become the complete standard in the training AI field.
On-device AI, based on edge AI semiconductors, is essentially the field where consumers can most tangibly feel the utility of AI technology. This is because the areas of application are endless, including person and object recognition based on computer vision technology, the Internet of Things (IoT), autonomous vehicles, AI electronics, and smart factories.
During a meeting with CHOSUNBIZ at the company’s headquarters in Pangyo on the 26th, Kim Nok-won, CEO of DeepX, explained, "If a data center made up of a huge number of GPUs is like being in the open ocean, then the edge computing field is more akin to the shallow waters close to land." He noted that the shallow sea requires smaller, more agile boats that can effectively avoid islands and coral reefs that may be unexpectedly confronted.
Kim added, "To analyze situations that the end-user's device experiences in real-time without delays using AI, chip solutions that can infer and make decisions independently without relying on data centers are necessary," and emphasized that "GPUs have long been proven unsuitable for this purpose. They are too heavy, expensive, and face serious issues concerning power consumption and heat generation, which is why neural processing units (NPUs) have emerged as the optimal solution."
DeepX, which has garnered attention as a promising player in the Korean AI semiconductor industry in recent years, will begin mass production of its first NPU product, DX-M1, based on a 5-nanometer process in the first half of next year. This chip, which is outsourced to Samsung Electronics’ foundry division for production, reportedly has a yield rate of nearly 90%. Unlike competing NPUs produced in the 10-20 nanometer range, DeepX aims to secure an overwhelming position in the market by producing NPUs using cutting-edge processes.
In particular, the DX-M1 to be launched in the first half of next year embodies the know-how that CEO Kim gained while working as a designer at leading semiconductor companies like Apple, Broadcom, Cisco, and IBM. By optimizing with software, it is expected to achieve higher performance while using less memory, and with a significant number of parameters optimized with cutting-edge processes, it should deliver at least four times the performance, power efficiency, and price competitiveness of chips currently on the market, while being only a quarter of their size.
Kim said, "Starting with this DX-M1, we are looking at a mid-to-long-term journey over the next decade, and DeepX aims to grow into a global fabless semiconductor design corporation valued at tens of trillions of won, not just a unicorn,” adding that "we will begin generating revenue from the second half of next year and embark on our journey as a full-fledged NPU corporation."
—If the last few years were a preparation phase, starting next year will be about verification in the global market.
"I think the product has come out at just the right time. If NPUs had come out in the 1990s, or even five years ago, they might have been ignored by the market. This is because there probably wouldn’t have been another ecosystem to fully implement AI technology on-device. Ultimately, the emergence and success of a new technology depend heavily on 'timing.' Until about three or four years ago, AI solutions in data center servers were not organized, but now there is a better understanding of the market and development. Just two or three years ago, we had to explain everything to our clients, but now they generally know how to use it. According to statistics from market research firms, it is expected that by 2026, 50% of IT hardware sold in China will be equipped with NPUs."
—AI is moving closer to users in the edge.
"Just looking at the applications the customers want, it is certain that NPUs will be equipped in all devices with cameras. This applies across the board from very small devices to large ones. In particular, the demand in the Chinese market is expected to be immense. The next market is robotics. Unmanned operations are now a complete trend in the industry. Automation is being implemented in many areas, including production, manufacturing, and logistics. We are in discussions for collaboration with many companies related to cameras and computer vision."
—There has been demand for edge computing for a long time, but global semiconductor corporations have not yet fully conquered it.
"The decision has already been made in the history of the semiconductor industry. Central processing units (CPUs) and graphic processing units (GPUs) are unsuitable for edge computing. This was confirmed as American semiconductor companies like Apple, Broadcom, and Cisco directly developed chips. That is the situation we are witnessing. There is no way to produce cheap edge chips through GPU-based designs from companies like NVIDIA, nor do they intend to. Intel is similarly situated."
From the perspective of corporate structural mechanisms, companies like NVIDIA and Intel cannot succeed in the edge market. The process of implementing on-device AI requires flexible design changes and shifts in real-time to fit the user environment. The shallow sea requires boats suitable for its depth. NVIDIA’s massive ships do not fit in these waters. For example, making design changes at a large company like NVIDIA incurs enormous expenses and resources. The operational costs do not align. With the same resources and time, DeepX can make 50 to 60 design adaptations with clients while NVIDIA can handle only one collaboration at a time. I believe this ecosystem will last for the next decade."
—Is DeepX's aim as a fabless company closer to Qualcomm or ARM than to NVIDIA?
"Yes. My goal is to be both Qualcomm and ARM in the era of AI semiconductors. There is now a consensus that on-device AI will be a game changer in the AI era. So, what chip will run on-device AI? A data center located far away won't be able to help with that. What AI technology must be processed immediately (without server connections)? That is why I rejected high salaries at global companies in Silicon Valley to start DeepX in Korea. NPUs are an indispensable presence for on-device AI and are the core of on-device AI."
—The DX-M1, which will be mass-produced next year, has quite different characteristics from existing NPUs.
"DeepX embodies a blend of Apple and Broadcom DNA. Looking at the size of the chip released this time, it is a quarter the size of mainstream NPUs available on the market, while reducing memory capacity, it has focused on optimization with software. The software-hardware co-optimization technology I learned while researching at Apple, Broadcom, and others is incorporated in it. By optimizing all optimal point parameters, the yield rate reaches about 90%."
—The most important marketing point is the performance and price approach towards clients. How has this standard been set?
"The 'criteria' in Silicon Valley is that there should be four times the performance and price advantage compared to existing products. That's where market separation begins. It is hard to change the market paradigm with just double the improvement. You need to win with at least quadruple the performance, and DeepX has prepared with this as our target. However, this is from the perspective of market prices, while companies that design and develop chips aim for ten or twenty times the performance. Setting those goals is necessary to achieve 4 to 10 times the performance and price competitiveness during the manufacturing and sales processes."
—There are predictions that you will go public (IPO) next year.
"There are still more things to show. A unicorn is a creature of imagination, but a unicorn corporation must be backed by actual results. DeepX's short-term goal is to contribute to creating a newly opening on-device AI market. An IPO is a tool to reach a company’s goal, not the goal in itself, so timing it appropriately in the process of fulfilling corporate objectives is ideal. Next year, since we've started mass production as DeepX, we have an obligation to generate revenue. I hope to go public when we can produce results worthy of public recognition and receive applause for our outcomes."