DingoBorat
Slim
The way this guy thinks, is a bit of a worry (in that he's good)..Hi Esky,
I agree and speaking of competition...
View attachment 85704
DEEPX aims to deliver energy-efficient, low-cost AI chip solutions for edge computing
News highlightsWednesday 28 May 20250
Exclusive Interview with DEEPX CEO Lokwon Kim. Credit: DEEPX
At Computex in Taipei, Lokwon Kim, founder and CEO of the Korean semiconductor company DEEPX, shed light on his company's robust capabilities in designing high-performance AI chips that prioritize cost-effectiveness and power efficiency. DEEPX aims to compensate, rather than compete with, industry giants like NVIDIA, by focusing on the burgeoning "on-device" AI market.
Kim came to sign an MOU of strategic partnership with Taiwan-based AIC Inc., a storage and server solutions provider. DEEPX's advanced Neural Processing Units (NPUs) will be integrated with AIC's robust industrial-grade server platforms. This collaboration aims to deliver unprecedented computational power combined with significant energy efficiency and compact form factors tailored specifically for edge environments.
Kim formalizes strategic partnership with Taiwan-based AIC Inc. through MOU signing. Credit: DEEPX
Kim's journey into AI chip design began during his PhD program at UCLA in 2007, where he had the opportunity to work on a deep learning processor project at the IBM T.J. Watson Research Center, which he calls "the number one research organization in the world." His research contributed to the early development of NPUs, even before the term "deep learning" was widely used.
"At that time there's no word like deep learning. We just call artificial neural network. So I was really extraordinary fortunate to start earlier than others," Kim shared. This early start gave him a significant advantage.
Having graduated from UCLA, Kim's tenure at Cisco, working on semiconductors for internet routers, provided crucial insights into the explosive growth of connected devices and the impending data deluge. He foresaw that by 2025, 70 billion devices would be connected to the internet, generating an unprecedented amount of data that humans alone could not process. Furthermore, he recognized that "40% of the internet data must be processed in a real time manner, not the data center we are waiting for." This realization fueled his vision for a low-power, high-performance, and low-cost AI processing solution at the device level.
"We need a very low power, high-performance, low-cost solution to process that on device, not the data center," Kim explained that relying on the cloud makes no sense due to latency, cost, and security problems.
DEEPX CEO Lokwon Kim at ex-Apple. Credit: DEEPX
He was inspired by the late Steve Jobs when he started working for Apple in 2014. "At the entrance of the Apple campus, there is this quotation from Steve Jobs that says, 'if you do something and it turns out pretty good, then you should go do something else that's wonderful…, Just figure out what's next." Kim decided it was time to answer that call when his internal startup idea was turned down by Apple. He decided to realize that dream by building a chip for the edge environment at his home country, South Korea, leveraging the foundry service of Samsung to support his fabless venture.
In 2018, Kim established DEEPX, and now the company has more than 100 employees in South Korea and will soon establish a branch office in July 2025, with FA engineer and salesperson stationed in Taipei.
"We Compensate the Giant, Not Compete"
Kim draws a parallel between the current AI chip landscape and the CPU market in the 1990s, where Intel dominated until ARM emerged with a more suitable solution for mobile devices due to its lower power consumption. He believes a similar dynamic is at play with AI processors, where NVIDIA's GPUs, while powerful, are not ideal for small, power-constrained devices due to their "high power consumption and very hot, high cost." As Kim put it, "We cannot put GPU solution into our small electrical devices. That's the point. So I wanted to solve it."
DEEPX's strategy is to address this gap by providing AI chips specifically designed for on-device applications, thereby "compensating Nvidia's market." This approach prioritizes:
• Real-time processing: "On device device zero latency guaranteed zero latency," Kim emphasized, highlighting the critical need for immediate responses in applications like self-driving cars and factory automation.
• Privacy and security: Kim explained the risk of sending factory data to data centers for AI processing, stating, "If you send the used data to the cloud for AI processing, hackers may intercept and leak that data. " On-device processing mitigates this risk.
• Total Cost of Ownership (TCO): DEEPX offers a significantly more affordable long-term solution. Kim provided an example: "The cost of our chip is under $100 for 10 years."
• Carbon emission reduction: Kim highlighted the environmental impact of current solutions: "Already, all the H100 GPUs in the world together consume more power than the total power consumption of France, which is one huge country." DEEPX's power-efficient chips offer a sustainable alternative.
DEEPX is already demonstrating its capabilities through collaborations with major companies. They are working with a large IT company in China on industrial monitoring and smart city projects. They are also partnering with South Korea company for white appliances, enabling functionalities like autonomous movement for robot vacuum cleaners and home security features such as detecting unauthorized individuals or elders' falling. DEEPX has already shared samples of their chips with over 300 global companies, achieving significant success in the pre-mass production market. Their chips are manufactured by Samsung's 5nm process.
DEEPX CEO Lokwon Kim at Computex Taipei 2025. Credit: DEEPX
DX-M2: Chip for Generative AI at the Edge
Looking ahead, DEEPX is developing DX-M2, their next-generation chip for generative AI. DEEPX aims to enable generative AI models with over 1 billion parameters, such as Meta LLaMA4 and Deepseek MoE, to run on-devices with just 5 watts of power consumption, and they are also considering TSMC 3nm or Samsung 2nm for future products.
Kim believes this will address the current financial challenges of generative AI. "Generative AI is not profitable right now because OpenAI pays huge energy bills to maintain its operations," said Kim, emphasizing DEEPX's solution would make generative AI accessible and affordable, with the chip costing under $50 and a module under $150, effectively eliminating the need for expensive data center charges for a decade once purchased. Kim believes this on-device generative AI will be a "hugely popular product." He acknowledges the skepticism, recalling similar disbelief when they announced their first chip's performance with low power, but he remains confident in their ability to deliver." Actually when we announced our DX-M1, nobody believed... But we did it. We proved it. Now we will do it again," he affirmed.
Lokwon Kim, CEO of DEEPX, on a mission to break into tech's top 10. Credit: DEEPX
Aiming to be Top 10 Players in Tech
DEEPX's long-term vision, as outlined by Kim, aligns with the suggestion of Jensen Huang, who he admires and considers a "rockstar." Huang's suggested phases for chip companies include:
1. Low cost and high usability: DEEPX has achieved this by creating chips that are highly useful and inexpensive.
2. Patent protection: DEEPX boasts over 300 patents for NPU technology, exceeding those of Intel, ARM, Qualcomm, and NVIDIA.
3. Ecosystem development: The next phase involves building an ecosystem with software frameworks and applications to increase the solution's value and profitability.
Kim's ultimate goal for DEEPX is to become "one of the major families in the world, within the top 10 players in the world" within the next 5 to 10 years, contributing to "the process of transformation process of human civilization, which is going to the super-intelligent."
This ambition, fueled by a decade of pioneering research and a clear strategic vision, positions DEEPX not just as a chip maker, but as a key enabler of a more intelligent, efficient, and sustainable technological future. As the demand for on-device AI continues to skyrocket, DEEPX's commitment to cost-effective, power-efficient, and high-performance solutions promises to democratize AI, making its transformative power accessible to countless devices and industries worldwide. To stay updated on the latest from DEEPX, follow the official DEEPX LinkedIn page.
But their tech doesn't seem "that flash" on the surface..
Their DX-M1 (and future DX-M2) are in 3nm process and using around 5w?..
That seems like a lot and the around $50 cost per chip, seems high too.. (possibly because of the process?)
AKIDA 2.0 IP in 3nm, would romp all over anything they currently, or intend to offer, in both performance and energy usage...
But the "guy" worries me..
), is also co-author of another (though similarly titled) paper describing research done with Akida: “Zero-Aware Regularization for Energy-Efficient Inference on Akida Neuromorphic Processor”, which happens to be presented at ISCAS (International Symposium on Circuits and Systems) 2025 in London tomorrow.
️
