Innoscience is advancing an all-GaN end-to-end power conversion technology within the NVIDIA MGX ecosystem to enable the next generation of high-density AI systems. Its 12kW 800V-to-48V design achieves approximately 99% peak efficiency and 98.2% full-load efficiency, while 150V GaN devices reduce the number of synchronous rectifiers by 50%. The solution supports the full range of intermediate bus voltages from 800V down to 48V, 12V, and 6V, and 15V GaN HEMTs enable high-frequency operation at 3 to 5 MHz, shrinking the size of magnetic components and capacitors. As AI workloads scale to rack-level and entire data centers, efficiency improvements in power semiconductor technology are breaking through rack power density limits, significantly reducing operational costs for high-performance computing facilities.
Author、Source: Wall Street Journal
As AI workloads scale to rack-level systems and entire data center sizes, power delivery has become a core bottleneck limiting data center system performance, density, and total cost of ownership. Within the NVIDIA MGX open modular reference architecture ecosystem, an efficiency revolution powered by all-gallium-nitride (All-GaN) technology is quietly reshaping the power delivery path from high-voltage distribution all the way to the GPU core.
The latest development in this technological evolution comes from Innoscience, a member of the NVIDIA MGX ecosystem. The company is advancing end-to-end All-GaN power conversion technology to support next-generation high-density AI systems. For investors and data center operators, this upgrade in underlying power semiconductor technology is critical to breaking through the upper limits of rack power density and significantly reducing the operational costs of high-performance computing facilities.
Traditional power delivery models are struggling to keep up with rising rack power demands; the challenge is no longer simply bringing power into the rack, but efficiently and compactly converting high-voltage electricity into the operating voltage required by GPUs. GaN technology, with its low on-resistance, low gate charge, and zero reverse recovery characteristics, is emerging as a key enabling solution, directly enabling smaller magnetic components, improved thermal performance, and lower total cost of ownership (TCO).
As AI systems move toward higher-density power architectures, the market is closely watching this power solution that breaks through physical and thermodynamic limitations. This will not only shorten the engineering and R&D cycle of accelerated computing systems but also significantly accelerate the large-scale commercialization of the next generation of AI factories.
Front-end conversion breakthrough: 12kW solution approaches 99% peak efficiency
As AI rack power consumption continues to rise, the front-end conversion stage has become one of the most demanding components in the power architecture.
In NVIDIA’s 800 VDC power architecture, reducing conversion stages by delivering DC power closer to the rack requires the front end to simultaneously handle high input voltage, high conversion ratios, and limited thermal budget and motherboard space.
Innoscience’s latest data demonstrates the direct benefits of GaN in this application. In its 12 kW 800 V to 48 V stage design, 650 V GaN double-sided cooled (DSC) devices are used on the primary side and 100 V GaN devices on the secondary side, achieving approximately 99% peak efficiency and 98.2% full-load efficiency at a switching frequency of 1 MHz. Additionally, the newly released 150 V GaN devices further simplify the secondary-side design, reducing the number of required synchronous rectifier devices by 50%. This reduction in footprint enabled by high-frequency operation delivers direct commercial value for AI systems pursuing higher rack density.
Beyond the 48 V front-end conversion, the choice of power architecture requires exceptional flexibility to meet diverse system design requirements for board space and thermal budget. Innoscience has expanded its All-GaN solution to cover the full range of intermediate bus voltage options, including 800 V to 48 V, 12 V, and 6 V.
For 800 V to 12 V conversion, the market can now leverage 40 V GaN devices to achieve efficient synchronous rectification and improved thermal performance; for 800 V to 6 V conversion, 15 V GaN devices as synchronous rectification solutions enable lower intermediate bus architectures, simplifying the final transition to GPU core voltage. In the critical 48 V to 12 V intermediate bus stage, Innoscience’s 100 V GaN solutions optimize multiphase buck conversion. Under the scale effects of AI factories, even minor efficiency gains translate to significant reductions in cooling requirements and operational costs.
Vertical power supply reshapes core response
In the final conversion stage closest to the computing core, traditional lateral power delivery faces significant challenges due to high current demands and critical transient response requirements, exacerbated by power distribution losses and complex motherboard routing. Vertical power delivery (VPD) is emerging as a viable architecture that offers shorter current paths, lower parasitic losses, and higher current density.
To meet the demands of rapid dynamic transients in GPUs, Innoscience has validated the feasibility of operating 15 V GaN HEMTs at frequencies between 3 MHz and 5 MHz, significantly reducing the required size of magnetic components and capacitors. The company is currently developing its DrGaN solution, which increases bandwidth by enabling high switching frequencies, thereby reducing reliance on traditional large-output capacitors. As future MGX AI systems continue to increase accelerator current density, power stages supporting VPD will become essential building blocks for near-die GPU power delivery.
To accelerate customer adoption, Innoscience offers a range of evaluation boards and reference designs to help system designers validate GaN performance across the entire AI power tree. These platforms include a 12 kW 800 V to 48 V demonstration board, a 4-phase 48 V to 12 V GaN evaluation board, and a 6 V DrGaN evaluation board designed for future vertical power delivery architectures.
The NVIDIA MGX ecosystem is driving the deployment of modular and scalable AI infrastructure. Amid growing power constraints on AI infrastructure, the evolution of power semiconductors must keep pace with increasing compute density. Through comprehensive coverage from 800 VDC down to GPU core voltage, more efficient and higher-density AI power infrastructure is accelerating from concept to reality.