MaxEpic, a nascent but rapidly advancing chiplet semiconductor startup based in Canada, is spearheading the development of next-generation fully-integrated power delivery and power integrity solutions crucial for the escalating demands of AI systems.
Founder and CEO Dr. Jerry Zhai is a seasoned veteran who has accumulated over two decades of experience in semiconductor technologies, product development, and business development with companies such as Analog Devices, Samsung Semiconductor and Renesas.
"AI has made a tremendous inroad in our industry from component to system level. The technology we have developed is a highly integrated power delivery chiplet that can address the new and rigorous demand of AI systems," said Zhai. "That is a fundamental force which drove us to start up this company."
The company, which began its focused development just about two years ago, is moving swiftly. Already, the company has successfully engaged with multiple major AI industry players, including large-scale hyperscalers, AI machine makers, semiconductor companies, and manufacturing partners, underscoring the immediate industry relevance of their solution.
Addressing the AI Power and Thermal Crisis
The dramatic advancements in AI hardware, such as GPUs, CPUs, and High Bandwidth Memory (HBM), have resulted in an unprecedented power dilemma. Dr. Zhai points out that the electrical current drawn by these components has increased three to five times over prior generations.
Modern CPUs, for instance, can consume up to 1,000 Amperes at a very low voltage of around 0.7 volts. The voltage for HBM continues todecrease. For example, the I/O voltage is down to 0.4V now to meet the low latency and high speed of data transmitting. The layer stack up of HBM that has increased up to 12 layers to meet workload demands has raised the HBM current consumption and density dramatically. This extreme current level and super high density create immense systemic challenges, including thermal dissipation issues and limits on I/O ports.
MaxEpic's innovative power delivery method trademarked as Ivy-ChipletTM is a fundamental solution to this challenge. It utilizes a proprietary IP to perform high-efficiency voltage conversion. Instead of drawing 0.7 volts directly, the device can effectively take a voltage that is twice or three times higher, such as 1.4 volts or 2.1 volts. In HBM I/O power which needs 0.4V, the voltage converstion ratio is 5-to-1 or 6-to-1 and the input voltage become 2.0V or 2.4V As a consequence of this voltage conversion, the current flowing into the device is instantly reduced by half, two-thirds or even less, depending on the designed voltage conversion . This significant reduction in current directly alleviates the problems of I/O port limit and thermal challengess by dramatically reducing power loss along the delivery path. The efficiency of MaxEpic's technology is remarkable, achieving up to 97% with its proprietary power streaming circuits
Technological Edge and Versatile Applications
The Chiplet's design offers significant flexibility in integration. It can be used as a standalone device or be seamlessly integrated directly into the AI device's package using contemporary 3D packaging technology.
MaxEpic's device is implemented using FinFET nodes, designed to be fully compatible with more advanced nodes like 7 nanometer, 5 nanometer, or even lower. This compatibility with cutting-edge process technology is vital, as the more advanced a chip is, the more efficient its power management needs to be.
The high efficiency, excellent thermal properties, and dimensional profile of MaxEpic's solution expand its potential beyond high-performance computing (HPC) and data centers. The technology can address systems requiring high degrees of integration and size reduction in three dimensions. Potential applications include next-generation AI glasses, which have a huge demand for high integration, small size and low profile, and Electric Vehicle (EV) systems, which benefit from the high reliability the technology offers.
Founder and CEO Dr. Jerry Zhai, MaxEpic. Credit: MaxEpic
Article edited by Joseph Tsai
