
Powering AI processors – 48V vertical power delivery changes the game
AI models demand huge current, driving legacy multiphase designs past 50 phases—causing space, thermal, and control limits Vicor addresses
Vicor 48V Factorized Power Architecture™ overcomes the current density and thermal limitations of traditional multiphase power architectures by separating voltage regulation from voltage transformation. By delivering high-voltage 48V directly to the point-of-load current multiplier, the system achieves a high current gain while minimizing conduction thermal losses across the power delivery network. This high-density, thermally efficient approach enables data centers to effectively meet the extreme low-voltage, high-current demands of next-generation AI processors.
By Maury Wood, VP of Strategic Marketing
Modern data centers face escalating power losses and efficiency bottlenecks when delivering massive levels of low-voltage current to high-performance processors driven by the explosion of AI workloads. In this interview, Maury Wood, Vice President of Strategic Marketing at Vicor discusses how 48V Factorized Power Architecture overcomes industry-standard multiphase TLVR conversion limitations, reducing conduction losses while meeting extreme multi-dimensional AI processor requirements. This innovative approach partitions voltage regulation and voltage transformation to mitigate accelerating thermal management challenges.
Tech Briefs: Today’s conventional power architectures are struggling to provide the high current density and low output voltage ripple required by next-generation AI processors. What are the technical implications?
Maury Wood: High volumetric current density, typically measured in A/mm3, requires the final-stage vertical power delivery (VPD) DC-DC converter to deliver high output current from a thin and small package. High-current-density VPD power modules enable direct mounting on the backside of the accelerator PCB where the bypass capacitors reside, simplifying assembly and improving thermal management and transient response.
TB: A top challenge for AI data centers is managing heat generated by the end-to-end power delivery chain. What is the best-practice approach to cut thermal losses across this chain?
Wood: Power is the product of voltage and current (P = IV). At a fixed power level, as voltage goes up, current goes down. Thermal losses accumulate from non-ideal DC-DC converter efficiency, and from conduction bus segments. DC-DC converter suppliers endeavor to improve continuous and peak conversion efficiency. Along conduction bus segments, sharp thermal losses arise from Ohm’s Law, (P = I2R), which dictates that resistive losses are proportional to the square of the distribution current. Powering an intermediate bus at the highest possible voltage and therefore lowest possible current for distribution across intermediate buses is therefore the clear winning strategy. The migration from 48VDC to 800VDC for rack power delivery is completely motivated by these simple physical laws. Similarly, bringing 48V as close as possible to the processor point-of-load is running in the right direction as compared to the 12V, 6V, and 2V intermediate bus voltages used with multiphase TLVR power architectures.
TB: How does the Vicor 48V Factorized Power Architecture fundamentally differ from today’s commonly used a 6V or 2V VR (IVR) multiphase approach for powering AI processors?
Wood: Multiphase TLVR power architectures use a fixed-ratio intermediate bus converter (IBC) to transform the nominal 48V / 50V input to the accelerator PCB to the required downstream regulator input voltage (6V or 2V). The downstream DC-DC converter regulates this input voltage to the sub-1V power domain rails of the chiplet-based AI processor. The Vicor 48V Factorized Power Architecture™(FPA™) reverses this functional partitioning. The first stage is regulation, in which the 50V input is tightly regulated to 48V with very low voltage ripple. The 48V regulator output feeds a current multiplier which transforms this voltage to sub-1V level needed by the AI processor. This 48V to 1V voltage transformation (48:1) or division correspondingly multiplies the current by the inverse factor (1:48), which Vicor calls “current gain.” Because 48V runs all the way to the final stage DC-DC converter (the current multiplier), the thermal losses along the intervening intermediate buses are dramatically reduced. For example, the thermal losses on a 6V intermediate bus are 64x higher than the thermal losses on a 48V intermediate bus.
TB: How much more effective is 48V Factorized Power Architecture compared to traditional 6V multiphase TLVR and 2V IVR architectures?
Wood: DC-DC converters, both fixed-ratio IBCs and voltage regulators operate with 90% to 95% continuous conversion efficiencies. Any future improvements in DC-DC converter efficiency will have relatively marginal impact on top-level AI factory power utilization effectiveness. On the other hand, power architectures which use relatively low-voltage intermediate buses, such as 6V or 2V, needlessly incur significant I2R thermal losses. These losses will become unacceptable as wafer-scale and panel-scale AI inferencing and training processors become mainstream over the next several years. Vicor sees 48V Factorized Power Architecture VPD using high current density, high current gain, and thermally-adept power modules as the only viable path forward as AI processors with 30kA and higher thermal design current inevitably emerge.
This article was originally published by Tech Briefs Magazine.
Maury Wood is Vice President of Strategic Marketing at Vicor Corporation. Prior to joining Vicor, Maury held senior roles at optical fiber test and semiconductor companies, including EXFO, AFL, Broadcom, NXP, Analog Devices, and Cypress. He holds a BSEE from the University of Michigan and did graduate studies at Northeastern University, Babson College, and MIT. He enjoys climbing, back-country skiing, mountain biking, and playing jazz bass.
Maury Wood, VP Strategic Marketing
Powering AI processors – 48V vertical power delivery changes the game
AI models demand huge current, driving legacy multiphase designs past 50 phases—causing space, thermal, and control limits Vicor addresses
Innovative 48V AI power architecture delivers both high current density and high current gain
Vicor 48V Factorized Power Architecture slashes thermal losses and delivers high current density to meet the extreme power demands of AI processors
48V Factorized Power Architecture™ slashes thermal losses and outperforms multiphase approach when powering AI
AI hardware faces severe power and cooling limits. Engineers debate 48V vs. 6V intermediate buses alongside new vertical power delivery (VPD) structures
大规模为人工智能(AI)供电
Maury Wood 在 Electronic Specifier 播客上介绍为人工智能(AI)供电所面临的挑战,并且揭示了晶圆级处理器的发展趋势


