PMICs roll out to support AI-enabled applications

Artificial intelligence (AI) applications require higher power and faster processing, which is driving some changes in the design of power management ICs (PMICs). These new requirements range from higher current support and better efficiency to improved thermal management and smaller solution size.

Power IC makers agree that AI capabilities are driving new requirements particularly for efficiency, transient response, and solution size. These requirements impact all market segments, whether it is for automotive, industrial, or data-center applications.

“The fast adoption of edge AI, i.e., endpoint devices that integrate AI capabilities, creates the need for PMICs that can power those devices with high efficiency to extend battery life, small solution size since most applications are size constrained, fast load transient capabilities, which doesn’t require additional capacitors in the application – so again, reduced solution size, and low EMI footprint since endpoint applications typically include RF and possibly, video and audio capabilities that are noise sensitive,” said James Lam, senior product marketing manager, Renesas Electronics Corp.

For automotive applications the need for higher power and faster processing are two key must-haves. “AI applications require significantly higher power and also require very fast response to delivering power compared to more conventional processing,” said Tom Sandoval, senior vice president, GM, Automotive Business Segment, Dialog Semiconductor. “Particularly for ADAS [advanced driver assistance systems] and AI, the “frame rates” are extremely high, this is where the individual snapshots of the situation being captured by the camera require faster processing and massively more of it.”

Sandoval said this has two effects on power management. “Number one, the individual cores in the SoC processing this data increase in numbers to parallel process the increased data and at a faster speed. This means the overall current requirements for the SoC are much higher. Number two, the cores increased processing speed means the current and voltage transition edges are much faster,” he said.

Dialog’s new DA914X-A multi-phase devices are optimized of this type of application. “They have better transient performance, lower losses, better efficiency, optimized thermal dissipation, and minimized ripple current and voltages than the competition’s single-phase equivalents,” Sandoval said.

Dialog designed the DA914X-A product family of high current, automotive-grade, step-down DC/DC (buck) converters for AI-based automotive applications. Targeting automotive AI SoCs, the DA914X-A family is an alternative to power solutions that require a combination of a power controller and discrete FETs. The new family integrates power FETs and all required control logic into a highly integrated monolithic device. Only a few external components are required for operation, which lowers the system bill-of-materials costs and solution footprints below 170 mm2, according to the company.

Sandoval said many customers use a set of discrete components but the DA914X family can meet these needs in a single monolithic PMIC with a smaller footprint, lower cost, and higher reliability.

“When compared to the competition, the DA9141-A has lower PCB costs and lower overall component heights for low-profile applications; has more flexibility, offering the opportunity to optimally place the Inductors and capacitors close to the point of load, and has a distributed power dissipation, resulting in a more even distribution of heat – critical for efficient thermal management,” he said.

The DA914X-A devices deliver current levels up to 40 amps, and are said to be extremely power efficient, reducing the thermal design challenges of powering complex automotive SoCs with very high current requirements. This makes the product family well-suited for powering graphics or AI embedded processors used in machine learning and vision applications for autonomous vehicles.

PMICs roll out to support AI-enabled applications

The DA914X-A product family for AI-based automotive applications (Source: Dialog Semiconductor)

The DA914X-A family currently consists of two devices: the DA9141-A and DA9142-A. The DA9141-A operates as a single-channel, quad-phase buck converter, delivering up to 40-A output current, while the DA9142-A operates as a single-channel, dual-phase buck converter, delivering up to 20-A output current. All devices have an input voltage range of 2.8 V to 5.5 V and an output voltage range of 0.3 V to 1.3 V, making them suitable for a variety of low power systems.

Key features of the DA914X-A products include multi-phase operation for better transient performance, lower losses, better efficiency, optimized thermal dissipation, and minimized ripple current and voltages when compared to a single-phase architecture. It also offers distributed power dissipation for more even distribution of heat and efficient thermal management. Other features include remote sensing, fully programmable soft start, and configurable GPIOs that includes support for I2C, DVC and Power Good indicator.

The family also offers dynamic voltage control that enables adaptive adjustment of the supply voltage dependent on the load, which increases efficiency when the downstream circuitry enters low power or idle mode, resulting in power savings, said the company.

The DA914X-A devices are AEC-Q100 Grade 1 qualified and are available in a
4.5 × 7.0 mm, 0.6-mm pitch 60-pin FC-BGA package. Industrial/commercial-grade versions are also available.

Renesas Electronics also offers a new highly integrated PMIC for AI processors, though aimed at industrial applications. The RAA215300 nine-channel PMIC complements Renesas’ RZ/V2L and RZ/G2L microprocessors (MPUs) designed for AI-enabled applications.

Lam said the new RAA215300 PMIC is well suited for edge AI applications. “It provides up to 15% higher efficiency than the competitor power solutions and the fast load transient response supports the the RZ MPU and DRP-AI applications without requiring additional output capacitors, and it also provides the complete DDR4 and DDR3L power solutions including VDDQ, VTT,  VREF and VPP supports, minimizing component count and solution size.”

In addition, the EMI and noise coupling reduction features simplify the RF, video, and audio sub-system designs targeted for high quality image and voice recognition AI products, he said.

The RAA215300 combines six buck regulators (supporting 5 A, 3.5 A, 2 × 1.5 A, 1 A , 0.6 A), three LDOs (supporting 2 × 300 mA, 50 mA), a real-time clock, and a coin cell/supercap charger. This level of high integration reduces design complexity and increases system reliability with fewer components on the board. It supports DDR4, DDR4L, DDR3 and DDR3L memory with dedicated VREF, VTT, and VPP rails. The PMIC also enables four-layer printed-circuit boards (PCBs), which reduce costs.

PMICs roll out to support AI-enabled applications

The RAA215300 nine-channel PMIC for industrial AI applications (Source: Renesas Electronics)

The device is configurable with a built-in EEPROM. The operating temperature range is -40°C to 105°C for industrial applications. Other features include spread spectrum to reduce EMI for RF applications, ultrasonic mode to eliminate audible noise coupling into microphones or speakers, and a built-in watchdog timer for secure system power up before any software is running..

The RAA215300 pairs with the RZ/G2L, RZ/V2L and several other Renesas offerings in a Winning Combination for a Scalable SMARC (Smart Mobility ARChitecture) system-on-module (SoM) with AI design. In addition to the Renesas MPU and PMIC, this Winning Combination includes power controllers, a USB PD controller, and a clock device.

Sample shipments of the RAA215300 are available now, and mass production is scheduled to start in the first quarter of 2022.

PMIC makers are also delivering new power ICs for communications and networking equipment, servers, and storage. One example is a multi-phase AI chipset from Maxim Integrated Products, Inc. This chipset powers AI hardware accelerators including GPUs, FPGAs, ASICs, and xPUs to increase solution efficiency and reduce solution size, said Steven Chen, director of business management for the Cloud and Data Center Business Unit at Maxim Integrated.

The MAX16602 AI cores dual-output voltage regulator and the MAX20790 smart power-stage IC deliver high efficiency and small total solution size for high-power AI systems. The chipset leverages the current ripple cancellation feature from Maxim’s patented coupled Inductor, claiming a one percent efficiency improvement compared to competitive solutions, translating into greater than 95 percent efficiency at 1.8-V output voltage and 200-A load conditions.

“There is an insatiable thirst for computing,” said Chen. “Power management ICs need to achieve high efficiency and support high current with a high-speed power management bus. New PMICs need to enable a customer’s AI solution by understanding the sweet spot of the power and feature requirements.”

Chen said there are several new requirements for PMICs, leading with the need for high current support (>500 A) with high efficiency, enhanced transient response, and low quiescent current. Other requirements include telemetry through PMBus, protection features, solution size to fit the form factor, cooling, and power delivery to optimize performance.

The new chipset hits on all of these points. The solution is scalable from 2 to 16 phases for different output current requirements (thermal design current is typically 60 A to 800 A or more), and the low-profile (<4 mm) coupled inductor is customizable to support multiple form factors such as peripheral component interconnect express (PCIe) and OCP accelerator modules (OAM).

AI systems implemented with the MAX16602 and MAX20790 multi-phase chipset are said to generate less heat compared to competitive solutions. Thanks to Maxim’s coupled inductor technology and monolithic integrated dual-side cooling power stage ICs, power loss is reduced as a result of a 50 percent lower switching frequency.

PMICs roll out to support AI-enabled applications

MAX16602CL8EVKIT board (Source: Maxim Integrated)

Maxim Integrated also noted that the monolithic integrated approach “practically eliminates the parasitic resistance and inductance between FETs and drivers” to achieve the high efficiency.

The MAX16602 with the coupled inductors and smart power-stage ICs implement high-efficiency core regulators with enhanced transient response and low quiescent current and the architecture helps reduce component count and enable advanced power management and telemetry and increase energy savings over the full load range. Regulator parameters for protection and shutdown can be set and monitored through the serial interface using the PMBus protocol.

The MAX16602 and MAX20790 devices are available at Maxim Integrated’s website (including samples) and its authorized distributors.

about Dialog Semiconductor
Maxim Integrated
Maxim Integrated Products
Renesas Electronics America