Kneron CEO on Business Strategy and Edge AI

AI Computing Power: A Shift from CPUs to MPUs - A Detailed Summary

Key Concepts:

• MPU (Memory Processing Unit): A new type of processor optimized for AI tasks, offering high performance with low power consumption, particularly suited for edge devices.
• GPU (Graphics Processing Unit): Traditionally used for graphics rendering, GPUs have become important for AI due to their parallel processing capabilities.
• CPU (Central Processing Unit): The traditional general-purpose processor, less efficient for AI workloads compared to GPUs and MPUs.
• HBM (High Bandwidth Memory): A high-performance RAM interface used in GPUs, currently less critical for MPU architectures.
• Edge Computing: Performing computation closer to the data source (e.g., on a device) rather than in a centralized data center.
• SMP (Symmetric Multiprocessing): A hardware architecture that allows multiple processors to access the same memory.
• NDA (Non-Disclosure Agreement): A contract that prevents the disclosure of confidential information.

1. The Evolution of AI Computing Hardware

The discussion centers on the shift in AI computing power from CPUs to GPUs and, most recently, MPUs. This evolution is likened to the progression from VHS to DVD to NBC (a small, high-capacity dongle), illustrating a trend towards increasingly compact and powerful processing units. While GPUs have been the dominant force in AI for over a decade, originating from their use in gaming and graphics, MPUs represent a new paradigm. MPUs, though currently small in scale, offer significant power for AI applications, surpassing CPUs in efficiency. The speaker emphasizes that the perception of CPUs being the best hardware for AI is inaccurate. The historical reliance on CPUs stems from their established infrastructure and the dominance of Google’s CPU-centric approach.

2. The Advantage of Edge Computing and MPUs

A key argument presented is the importance of competitive advantage through edge computing. As hyperscalers focus on centralizing data in data centers, the ability to process data locally on edge devices becomes crucial. The company highlighted has been pioneering SMP technology for ten years, even publishing the first SMP architecture and facing challenges as a result. The speaker suggests that their published work serves as a foundational textbook for the field. MPUs are particularly well-suited for edge computing due to their small size and low power consumption, making them ideal for applications like wearable devices where CPUs are impractical due to heat and speed limitations.

3. MPU Architecture and its Benefits

The MPU architecture is described as being structurally different from GPUs and CPUs, akin to a small PC Tango (presumably a reference to a compact computing device). Unlike GPUs, which heavily rely on memory technologies like HBM and 3D packaging, MPUs are less demanding in this regard. This difference in architecture allows for greater efficiency and potentially lower costs. The speaker highlights the potential for businesses to leverage MPUs to build their own AI applications, reducing reliance on external providers like OpenAI. This is achieved through the creation of small, on-device APIs.

4. Real-World Applications and Partnerships

The company has established partnerships with major players in various industries. Notable customers include Naver, the largest search engine in Korea, and a significant, unnamed US customer (under NDA). They also work with Foxconn and Qualcomm. Geographically, their primary markets are Korea, Japan, the US, and the Middle East (with a new site in Saudi Arabia). They also have a presence in China, primarily in consumer applications like AI-powered air conditioners and sensitive areas like military defense.

5. Competitive Landscape and Innovation

The speaker acknowledges the increasing competition, particularly from China, which views AI as a key area of geopolitical competition. China is described as being “very aggressive” in its pursuit of AI self-sufficiency. However, the speaker maintains that core 0 to 1 innovation still resides on the Western side. They also position their company as a pioneer in the field, with many competitors essentially building upon the foundations laid in their published work (“We can write a textbook means they received writing in the textbook. It's in their own pattern.”). Interestingly, they also collaborate with competitors like NVIDIA, with Qualcomm acting as both an investor and an IP partner.

6. Future Outlook and IPO Plans

The company anticipates strong growth driven by the increasing demand for on-device AI. They believe their technology will enable users to create “personal GTPs” (likely referring to Generative Pre-trained Transformers) with high privacy and low latency. Regarding future plans, the company is preparing for an IPO, aiming for a “meaningful” listing with a solid customer base and industry influence. They are awaiting a significant contract (expected in Q2 of this year) to demonstrate their market leadership. The possibility of being an acquisition target is acknowledged, but the primary focus is on achieving a successful IPO.

7. Data and Statistics

• The company’s major income sources are Korea, Japan, the US, and the Middle East.
• They have a customer that is the largest AI conditioner company globally.
• They are working with a global top five customer whose market share is significant.
• Qualcomm is an investor and IP partner.

Synthesis/Conclusion:

The conversation highlights a significant shift in the AI computing landscape, moving away from traditional CPU and GPU dominance towards the more efficient and versatile MPU architecture. This shift is particularly important for edge computing, enabling localized AI processing with reduced power consumption and latency. The company discussed is positioned as a pioneer in this space, with established partnerships and a clear roadmap for future growth, including a planned IPO. The increasing competition, especially from China, is acknowledged, but the speaker remains confident in the company’s innovative edge and its potential to shape the future of AI.