AI accelerator cards

Design efficient AI accelerator cards and high-current AI chips using Infineon’s low-voltage DC-DC processor power solutions for heavy AI workloads

nobutton

Overview

Infineon's low-voltage DC-DC processor power solutions for AI accelerator cards provide high power conversion efficiency, fast transient response, and precise telemetry. This lets you design AI chips with superior system performance and optimized resource utilization under heavy AI workloads. By utilizing advanced TLVR topologies and integrated modules, Infineon solutions enhance thermal management and reduce board footprint for modern AI accelerator cards.

Benefits

  • Peak power conversion efficiency
  • Fast transient response
  • Accurate telemetry reporting
  • Lower capacitor count with TLVR
  • Long-term system reliability
  • Space-efficient power designs
  • Industry-leading performance
  • Reduced TCO for AI data centers

About

AI accelerator cards are specialized hardware expansion cards designed to accelerate complex generative AI and machine learning workloads at scale. 

What is the difference between AI accelerator cards and GPUs?

Unlike general-purpose GPUs, which support a range of graphics and compute tasks, AI accelerator cards are optimized to enable greater energy efficiency and throughput for AI workloads and neural network operations such as tensor processing and high-throughput inference.

How does Infineon power next-gen AI accelerator cards?

Infineon provides a complete power architecture for AI accelerator card applications, combining 48 V power distribution solutions with common-footprint XDP™ digital multiphase controllers, OptiMOS™ high-density dual-phase power modules, and OptiMOS™ power stages.

These technologies support trans-inductor voltage regulator (TLVR) topologies, enabling industry-leading power delivery to high-current AI chips and SoCs while reducing output capacitance and system cost, and saving PCB space.

Together, these solutions enable efficient, scalable power delivery networks for AI accelerator cards used in hyperscale data centers and AI compute platforms.

Power delivery network (PDN) is the infrastructure that delivers power from the voltage regulator to the CPU and other components on the server motherboard. It is critical for supplying stable voltage to high-current AI processors used in AI accelerator cards. 

Modern AI chips generate extremely fast current transients and high di/dt load steps, making it difficult for traditional multiphase buck converters to maintain tight voltage deviation limits.

How to mitigate PDN challenges in AI accelerator cards?

The trans-inductor voltage regulator (TLVR) architecture addresses this challenge by using transformers with a 1:1 turns ratio instead of traditional inductors. This approach improves transient response and enables better voltage regulation without needing you to place the multiphase voltage regulator closer to the processor.

Infineon’s TLVR solutions help mitigate PDN limitations in multiphase DC-DC processor power systems, enabling reliable voltage regulation for AI processors requiring 200 A to 1000 A current with highly dynamic load behavior.

Discover how Infineon’s TLVR solution improves power delivery performance in AI accelerator cards.

AI accelerator cards need high-current power delivery to support demanding AI workloads. Hence, power solutions must provide high efficiency, current density, and compact design while minimizing thermal losses.

How to address AI accelerator performance challenges?

Infineon addresses these challenges with its OptiMOS™ dual-phase power modules, including TDM2354xD and TDM2354xT. These 160 A modules integrate two OptiMOS™ 6 power stages, inductors, and capacitors on a single substrate, enabling two phases of a multiphase buck regulator in a compact footprint 40% smaller than equivalent discrete solutions.

How to handle high power demands?

For higher power requirements, OptiMOS™ TDM2454xx quad-phase modules deliver peak currents up to 280 A, proprietary magnetics, exceptional power density of 2.0 A/mm² in a compact package optimized for vertical power delivery (VPD) architectures.

How to ensure thermal performance?

The modules feature Infineon’s proprietary inductor-on-top design that improves thermal performance while supporting high-density AI workloads. Also, Infineon's robust OptiMOS™ 6 trench technology and chip-embedded package enhance electrical and thermal efficiencies.

AI accelerator cards are specialized hardware expansion cards designed to accelerate complex generative AI and machine learning workloads at scale. 

What is the difference between AI accelerator cards and GPUs?

Unlike general-purpose GPUs, which support a range of graphics and compute tasks, AI accelerator cards are optimized to enable greater energy efficiency and throughput for AI workloads and neural network operations such as tensor processing and high-throughput inference.

How does Infineon power next-gen AI accelerator cards?

Infineon provides a complete power architecture for AI accelerator card applications, combining 48 V power distribution solutions with common-footprint XDP™ digital multiphase controllers, OptiMOS™ high-density dual-phase power modules, and OptiMOS™ power stages.

These technologies support trans-inductor voltage regulator (TLVR) topologies, enabling industry-leading power delivery to high-current AI chips and SoCs while reducing output capacitance and system cost, and saving PCB space.

Together, these solutions enable efficient, scalable power delivery networks for AI accelerator cards used in hyperscale data centers and AI compute platforms.

Power delivery network (PDN) is the infrastructure that delivers power from the voltage regulator to the CPU and other components on the server motherboard. It is critical for supplying stable voltage to high-current AI processors used in AI accelerator cards. 

Modern AI chips generate extremely fast current transients and high di/dt load steps, making it difficult for traditional multiphase buck converters to maintain tight voltage deviation limits.

How to mitigate PDN challenges in AI accelerator cards?

The trans-inductor voltage regulator (TLVR) architecture addresses this challenge by using transformers with a 1:1 turns ratio instead of traditional inductors. This approach improves transient response and enables better voltage regulation without needing you to place the multiphase voltage regulator closer to the processor.

Infineon’s TLVR solutions help mitigate PDN limitations in multiphase DC-DC processor power systems, enabling reliable voltage regulation for AI processors requiring 200 A to 1000 A current with highly dynamic load behavior.

Discover how Infineon’s TLVR solution improves power delivery performance in AI accelerator cards.

AI accelerator cards need high-current power delivery to support demanding AI workloads. Hence, power solutions must provide high efficiency, current density, and compact design while minimizing thermal losses.

How to address AI accelerator performance challenges?

Infineon addresses these challenges with its OptiMOS™ dual-phase power modules, including TDM2354xD and TDM2354xT. These 160 A modules integrate two OptiMOS™ 6 power stages, inductors, and capacitors on a single substrate, enabling two phases of a multiphase buck regulator in a compact footprint 40% smaller than equivalent discrete solutions.

How to handle high power demands?

For higher power requirements, OptiMOS™ TDM2454xx quad-phase modules deliver peak currents up to 280 A, proprietary magnetics, exceptional power density of 2.0 A/mm² in a compact package optimized for vertical power delivery (VPD) architectures.

How to ensure thermal performance?

The modules feature Infineon’s proprietary inductor-on-top design that improves thermal performance while supporting high-density AI workloads. Also, Infineon's robust OptiMOS™ 6 trench technology and chip-embedded package enhance electrical and thermal efficiencies.

Documents

Design resources

Developer community