POL Converters Feed Proliferating Power Rails in AI Chips

As next-generation AI chips move into the multi-kilowatt range, power architectures are being pushed to their limits. The core power rails in chips such as NVIDIA's Blackwell GPU require extreme currents of several thousand amps and ultra-fast transient response times, and future platforms like NVIDIA's Rubin GPU are expected to push these demands even further.

At the same time, the number of non-core rails is increasing rapidly, each requiring compact, efficient regulation with precise monitoring and control.

At APEC last month, Infineon introduced a family of fully integrated DC-DC buck regulators aimed at these non-core rails: the TDA49720, TDA49712, and TDA49706. The devices complement the company’s higher-current multiphase voltage regulators. They're targeted at the core rails in GPUs and other processors such as Arm's chiplet-based AGI CPU that support them during AI training and inference. The all-in-one power modules are capable of delivering power vertically up through the circuit board.

The point-of-load (POL) converters can deliver 6, 12, and 20 A of current in compact 3- × 3-mm and 3- × 3.5-mm packages. Infineon said the small footprint allows placement close to the load, improving power density and simplifying layout within densely packed accelerator cards and server boards. The devices operate over a 2.7- to 16-V input range and across junction temperatures of –40 to 150°C.

The chips also come with PMBus-compliant digital telemetry. This enables accurate reporting of important parameters, including output voltage, load current, input voltage, and die temperature, leading to system-level optimization and reliability monitoring.

Infineon said a proprietary valley current-mode, constant on-time control scheme enables fast transient response and cycle-by-cycle current limiting. That also helps enable stable operation required by all-MLCC output-capacitance designs.