Founded in 2018, TetraMem is delivering disruptive computing fabric based on in-memory computing (IMC) technology for AI applications.
After a long career at HP, why have you decided to put all your heart and energy into TetraMem?
It was not an easy decision to leave HP, a company I greatly respected. However, I chose to resign from my long tenure there and co-found TetraMem with three other co-founders because we recognized a critical need to advance efficient AI computing, especially as AI has become increasingly pervasive since AlphaGo's historic victory over Lee Sedol in 2016. Technologies like ChatGPT, Sora, and Apple's Vision Pro are groundbreaking, but we're still just beginning to realize their full potential. The demand for AI computing power and efficient data centers is skyrocketing, yet we’re hitting significant efficiency limits with current technology. The traditional digital processor approach may never reach Artificial General Intelligence (AGI).
Industry leaders like Dr. Lisa Su from AMD have emphasized the massive growth in data center usage and the associated challenges in power consumption and computing efficiency. As data generation surges and AI algorithms grow more complex, a hardware bottleneck looms that could impede further progress. TetraMem’s technology is designed to directly address these challenges, paving the way for more sustainable and efficient computing, which is essential for the future of AI and beyond. This is about commoditizing AI as if it is water or electricity. d
How is your pioneering analog in-memory computing (AIMC) addressing this clear need in the market?
The math of AI primarily revolves around matrix operations, which requires extensive data movement between the processor and memory in today's digital processors.Inspired by the efficiency of the human brain, our solution addresses fundamental challenges in current systems, specifically the memory wall, heat wall and end of Moore’s scaling. While in-memory computing, photonics computing, and quantum computing are seen as potential solutions, quantum and photonics computing face significant technical challenges and are not suitable for AI applications. In contrast, in-memory computing minimizes data movement by integrating memory and computation in the same physical location within a crossbar-array structure, which naturally aligns with matrix operations.
Our approach creates an in-memory computing unit where the non-volatile memory stores neural network parameters via conductance and performs computations through physical laws, achieving brain-level efficiency. By applying Ohm's law—where current equals voltage times conductance—we handle multiplication. By collecting the current across multiple rows or columns of the crossbar, we achieve accumulation. This method allows us to perform large matrix operations efficiently in a single cycle. In comparison, digital approaches require millions of transistors and multiple clock cycles to perform even an 8x8 matrix operation, consuming significantly more energy. Our technology is not just theoretical. We have already developed the MX100, the world's first 8-bit multi-level RRAM in-memory computing evaluation kit.
How can TetraMem achieve what larger companies, like Intel or Nvidia, might not?
TetraMem is able to achieve what larger companies like Intel or Nvidia might not due to our strategic focus and agility. While big companies often must prioritize short-term returns and established product lines, TetraMem can dedicate itself entirely to pioneering breakthroughs in neuromorphic computing. Our technology addresses the inefficiencies of traditional digital systems by leveraging the potential of RRAM (Resistive Random-Access Memory) for computational applications—an area that remains underexplored by major players.
Our roots trace back to fundamental research, where we identified the transformative potential of RRAM beyond just memory storage, extending its use to computation. This foundational insight led to the creation of TetraMem in 2018, allowing us to explore and develop solutions with a depth and focus not often possible in larger organizations. At TetraMem, we possess deep expertise across six critical dimensions: device modeling, process development, material science, application design, algorithm optimization, and circuit design. This comprehensive understanding allows us to optimize memristor-based devices specifically for neuromorphic computing, setting us apart from larger companies that may not have the same level of specialized focus or flexibility.
Our ability to innovate in this space, unencumbered by the strategic constraints of larger corporations, enables us to push the boundaries of what is possible in AI hardware.
Could you explain the three waves of computing - GPUs, in-memory computing and neuromorphic computing - to our readers?
The first wave of computing has been dominated by GPUs and other digital ASICS like TPU, especially in data centers for AI training and inference. While GPUs and TPUs have been crucial in advancing AI, they face significant challenges related to energy efficiency, with throughput increasingly constrained by the limits of available electricity and cooling resources.
As we move forward in the very near future, we're entering the second wave, characterized by analog in-memory computing. This approach offers a more energy-efficient alternative by integrating computation directly within memory devices, reducing the need for data movement and thereby improving overall efficiency.
The third wave looks to neuromorphic computing, which seeks to emulate the structure and function of the human brain by using spiking neurons for processing. This method aligns more closely with how our brains operate, promising significant advancements in both efficiency and capability, potentially enabling a new generation of AI applications that are more powerful and energy-efficient than ever before. However, for neuromorphic computing, we are still in the early research stage. Both device and software sides need years of research before we can see the light of the real applications.
In your crystal ball, when do you think Artificial General Intelligence (AGI) will be achieved?
Predictions for achieving Artificial General Intelligence (AGI) vary widely, with some sources, like Time magazine, suggesting it could happen by 2045. However, much like you can't send people to the moon with a steam engine, achieving affordable AGI with purely digital approaches may be unrealistic. Our work is focused on advancing analog in-memory computing, which we believe is a crucial step toward AGI.
Unlike traditional digital computing, analog computing, particularly with the use of memristors, more closely mimics how the brain processes information. This alignment with natural neural processes could accelerate the development of AGI. Recent studies indicate that significant breakthroughs in analog computing technology could bring us closer to AGI as early as 2035, potentially offering a faster and more efficient path to this milestone.
What are some of your key challenges that you anticipate facing in the coming few years?
As a small startup, we anticipate facing significant challenges and opportunities in both engineering and commercialization over the next few years. While our technology has achieved recognition through prestigious publications in journals like Nature and Science, translating these innovations into market-ready products presents a major hurdle. We must navigate complex engineering challenges to bring our advanced technologies into practical, scalable applications that meet market demands.
Establishing a strong market presence amidst intense competition from established digital solutions is also critical. The transition from digital to analog computing will not happen overnight—just as it took decades for transistors to replace vacuum tubes, we expect a similar timeline for analog computing to gain substantial market share in AI applications. Our focus will be on overcoming these obstacles while driving the adoption of our technologies in a rapidly evolving industry.
Where can we expect TetraMem to be in three years, and what sort of funding will you need to get there?
Though the U.S. semiconductor and chip startupecosystem has been picking up,s a small startup, we still anticipate facing significant challenges and opportunities in both engineering and commercialization over the next few years. While our technology has achieved recognition through prestigious publications in journals like Nature and Science, translating these innovations into market-ready products presents a major hurdle. We must navigate complex engineering challenges to bring our advanced technologies into practical, scalable applications that meet market demands.
Establishing a strong market presence amidst intense competition from established digital solutions is also critical. The transition from digital to analog computing will not happen overnight—just as it took decades for transistors to replace vacuum tubes, we expect a similar timeline for analog computing to gain substantial market share in AI applications. Our focus will be on overcoming these obstacles while driving the adoption of our technologies in a rapidly evolving industry.
A final note, could you explain your slogan - 'make processors cooler, make machines warmer' - as if I were a ten-year-old?
Our slogan, "make processors cooler, make machines warmer," reflects our mission to develop more powerful processors that drive not only artificial intelligence but also the future of emotional intelligence. This approach helps robots and machines become more helpful and human-like. As Elon Musk has pointed out, our world is designed by, around and for humans, so robots must integrate seamlessly and work alongside us. We refer to this as embodied AI, where machines are crafted to be a natural part of our daily lives.
However, we aim for more than just smart robots; we want them to possess qualities like empathy and compassion, making them feel "warmer"—almost as if they have a heart. Achieving this level of emotional intelligence is both crucial and challenging, requiring advanced processors and more sophisticated AI neural networks. In our logo, "E" stands for Emotional Intelligence and "A" for Artificial Intelligence, highlighting this dual focus.
