[Editor's Note] As power demand in data centers skyrockets due to the spread of AI, the existing 48V architecture has reached its limits, and a transition to an 800V DC-based structure is required. The transition to 800V DC necessitates the redesign of the entire power system, including PSUs, IBCs, and BBUs, and wide-bandgap technologies such as GaN and SST are emerging as key factors in realizing high efficiency, high density, and high reliability. In particular, as we enter the era of 1MW per rack, power, cooling, and safety requirements are rapidly increasing, and it is being discussed that an approach viewing the entire system as a single power chain is necessary for design. Accordingly, we arranged a meeting with Pradeep Shenoy, a data center power product expert and Lead in Computing Power Technology at Texas Instruments (TI), to hear about product development and related solutions for data center power.


■ Please briefly introduce yourself and explain your current role in the data center power sector at Texas Instruments.

As a Compute Power Technologist in the TI Data Center Systems Engineering team, he leads the definition and development of power electronics solutions for TI's data center portfolio.

Specifically, he/she is responsible for designing high-efficiency, high-density converter architectures, verifying silicon and system performance, and collaborating with customers and internal product groups to ensure that TI's data center power products meet the stringent efficiency, reliability, and form factor requirements of modern hyperscale and enterprise servers.

■ He stated that a fundamental redesign of the existing power architecture is necessary. I wonder why the existing method alone is no longer sufficient.

Data centers are the foundation of the information age.

And the explosive growth of AI and machine learning is fundamentally changing the way data centers are powered.

The core problem is that computing power density is increasing rapidly at all system stack levels.

Until very recently, the power consumption of processors was in the hundreds of watts (W).

On the other hand, individual AI processors today consume power in kilowatt (kW) units.

Looking at it on a rack-by-rack basis, the rated power, which was at the level of tens of kW, is rapidly exceeding 100 kW, and will soon exceed 1 MW per rack.

At the overall data center level, power consumption, which was tens of megawatts, has already grown to hundreds of megawatts, and future AI data centers will require more than gigawatts (GW) of power.

Existing 48V-based architectures are not designed to handle loads of this scale.

Traditional approaches have already reached their limits, and we are now facing a major shift.

That is precisely why a comprehensive system-level redesign of the power architecture is a necessity, not an option.

■ What are the key advantages of the 800V architecture compared to existing 48V-based systems? In addition, the most important technical challenge that must be solved when introducing 800V DC is

The transition to 800V frees up physical space within the IT rack by moving AC/DC converters and battery backup units (BBUs) to a sidecar power rack separated from the IT rack, thereby increasing computing density.

The 800V busbar is positioned along the rear of the rack, eliminating the need for an intermediate DC/DC power shelf.

However, this transition entails significant technical challenges.

The PSU (power supply unit) must be expanded from the existing 8kW class to the 30kW class and must receive a 3-phase AC input and output 800V or ±400V DC.

To achieve this, new circuit topologies such as 3-level flying capacitor PFC (power factor correction) and delta-delta connected 3-phase LLC stages are required.

Battery backup and capacitor bank devices must also be redesigned to suit high-voltage DC environments.

The intermediate bus converter (IBC) must step down the voltage from 800V to a high conversion ratio (e.g., 64:1 or 128:1) while maintaining efficiency and reliability during this process.

Since 800V is applied directly to the IT tray, safety requirements must be incorporated from the system design stage.

Finally, robustness and reliability are factors that can never be compromised.

No one can afford a situation where the failure of a single relatively inexpensive power component brings down an entire expensive computing rack.

■ When do you expect the transition from Gen3 to Gen4 (Solid State Transformer-based Architecture) to become a reality?

3rd generation architecture (Gen3) is IT rack computing While excellent for improving density, sidecar power racks are still required to occupy data center IT floor space.

Therefore, the next step in the evolution of data centers is to remove the AC/DC power conversion function of the sidecar and distribute 800V directly to IT racks via the busway.

The power distribution method for each generation requires numerous sophisticated power conversion functions.

These functions include PFC, DC/DC conversion of 800V DC or ±400V DC, diode OR-ing, current sharing, hot swap, protection circuits, control, power metering, etc.

Advanced semiconductor technology is key for all these functions to operate with maximum performance and efficiency.

A Solid State Transformer (SST) integrates AC/DC and DC/DC conversion into a single high-frequency stage.

Because it operates at a much higher frequency than conventional commercial frequency transformers, its physical size is significantly reduced, and it enables direct distribution of 800V (or higher) DC busways within data centers.

It also supports close integration with on-site energy resources, such as solar power generation and energy storage systems, which is becoming an increasingly important factor in data centers improving energy efficiency.

■ Why has GaN emerged as an essential technology for data center power innovation? How do you see GaN technology developing in the future, particularly in terms of expanding the voltage range?

GaN (Gallium Nitride) is a wide-bandgap semiconductor that delivers performance superior to conventional silicon devices in high-voltage and high-frequency power conversion applications.

As data centers transition to 800V DC power distribution and move toward 1MW per rack, the switching frequency, efficiency, and power density requirements for power conversion hardware are exceeding the levels that silicon can reliably handle.

GaN is the material that makes this next-generation system possible.

TI's approach does not stop at simply providing GaN transistors.

The key to fully realizing the potential of GaN lies in integration.

TI reduced the common source inductance to less than 1 nH and the gate loop inductance to less than 4 nH by directly integrating the gate driver within the same package as the GaN FET.

This holds great significance.

This is because at high switching frequencies, even very small parasitic inductance significantly increases switching losses and poses a reliability risk.

Driver integration reduces these risks and enables customers to drive at switching frequencies where they can actually achieve the power density and efficiency benefits of GaN.

Simply put, using GaN without an integrated driver is like driving a sports car only in the low-speed lane.

It amounts to having the hardware but failing to properly utilize its performance.

The strategic goal is clear.

GaN enables system cost reduction, increased power density, and simplified design, and TI’s vertically integrated approach, ranging from process to packaging and driver integration, supports customers in practically realizing these benefits in production systems.

■ If the deployment of 1MW per rack becomes a reality, what do you expect to be the biggest changes in terms of power, cooling, and safety?

From a power architecture perspective, to reach 1 MW per rack, a complete transition from 48 V distribution to an 800 V DC busbar architecture is essential, as described in Gen3 and Gen4.

At this scale, efficiency becomes even more important.

A 1% point conversion loss is converted into tens of kW of heat, which not only increases the amount of heat that needs to be managed but also represents wasted power that does not contribute to computing.

Therefore, the evolution of architecture that reduces conversion steps and increases efficiency is not merely a matter of power supply, but is directly linked to cooling costs and the overall operational economics of the data center.

Safety also presents a fundamentally different level of challenge compared to the 48V environment.

High voltage must be carefully designed from the outset for every element of the system, from the power distribution boards of IT trays to busbar infrastructure and circuit protection devices.

Liquid cooling CDUs (Coolant Distribution Units) are gradually evolving into intelligent systems.

High-output pumps, precise sensor-based control, and enhanced reliability are required to handle the increasing thermal loads of AI data centers.

TI is at the forefront of this trend.

TI's extensive power and signal The chain and embedded product portfolio provides a one-stop solution for CDU electronic design, covering all areas ranging from control and drive, sensing and safety, to power supply for high-performance liquid-cooled servers.

■ The most important message I want to emphasize to data center power designers or IT infrastructure companies at this point is

Data centers are the foundation of the information age in which we live, and TI provides core solutions across all areas of data centers, from power architecture to cooling and networking.

The transition to 800V DC power distribution means a fundamental redesign of the data center power architecture.

It is not a simple upgrade, but a complete redesign.

And the combination of this architectural evolution and advanced wide-bandgap semiconductor technology such as GaN is the factor that makes it possible to realize more than 1 MW per rack.

The growth of AI infrastructure over the next 10 years is impossible without this level of innovative transformation.

My advice to the designers and infrastructure teams embarking on this journey is to start designing by viewing the entire power chain from a single perspective, from AC grid connections to processor core voltages.

Decisions made at each stage interact with and influence all other stages.

It is precisely from this system-level perspective that substantial results can be achieved in terms of efficiency, power density, reliability, and cost.