NVIDIA leads the charge in the 800V DC field, presenting new opportunities for power chip manufacturers.
At the OCP Global Summit, NVIDIA focused on the future development of gigawatt-scale AI factories and presented a series of cutting-edge technologies and innovative achievements. Among them, the 800V direct current (VDC) technology emerged as a major highlight, leading the transformation of the data center energy architecture.
Compared with the traditional 415 or 480V alternating current (VAC) three-phase systems, the 800V DC architecture demonstrates significant advantages. From a physical transmission perspective, the same copper cable can transmit over 150% more power under 800V DC. The 200-kilogram copper busbars previously required to power a single rack can be significantly reduced, saving customers millions of dollars in costs.
In the practical application of data centers, the 800V DC architecture enhances the system's scalability, enabling data centers to easily handle the ever-growing computing power requirements. Its higher energy efficiency reduces power transmission losses, aligning with the current trend of green energy conservation. At the same time, it reduces material usage, optimizes the cost structure, and brings higher performance capacity to data centers. In fact, the electric vehicle and solar industries have long adopted 800V DC infrastructure for similar benefits, and now the data center field is also embracing this wave of transformation.
Foxconn actively responded by announcing the 40-megawatt Taiwan Kaohsiung No. 1 Data Center built for 800V DC. More than 20 industry pioneers, including CoreWeave and Lambda, have also joined the design of 800V DC data centers. Additionally, Vertiv launched the space-, cost-, and energy-saving 800V DC MGX reference architecture, and HP announced its support for related technologies, jointly improving the 800V DC ecosystem.
More than 20 NVIDIA partners are helping to provide rack servers that meet open standards, contributing to the future gigawatt-scale AI factories.
Chip Providers: Analog Devices, Inc. (ADI), AOS, EPC (Efficient Power Conversion), Infineon, Innoscience, MPS (Monolithic Power Systems), Navitas Semiconductor, onsemi, Power Integrations, Renesas, Richtek, ROHM, STMicroelectronics, and Texas Instruments
Power System Component Providers: BizLink, Delta, Flex, GE Vernova, Lead Wealth, LITEON, and Megmeet
Data Center Power System Providers: ABB, Eaton, GE Vernova, Heron Power, Hitachi Energy, Mitsubishi Electric, Schneider Electric, Siemens, and Vertiv.
Among them, there are quite a few partners from the Chinese mainland and Taiwan region. In particular, Innoscience has become the only partner in the local chip industry. Additionally, companies such as PI have newly joined the ecosystem.
Innoscience - The Only Gallium Nitride IDM
As the only full-stack gallium nitride supplier and a leading gallium nitride IDM enterprise in the industry, Innoscience is the only company that has achieved mass production of gallium nitride from 1200V to 15V and can provide full-link solutions from 800V to 1V. This makes Innoscience the only supplier capable of providing full GaN power solutions for all conversion stages, easily coping with the evolution of future architectures to meet higher power requirements.
Innoscience officially stated that traditional artificial intelligence systems based on a 48V voltage are facing severe challenges - low efficiency and high copper consumption, with more than 45% of the total power consumption being wasted on heat dissipation. If future artificial intelligence clusters (such as racks equipped with more than 500 GPUs) continue to use the old PSU power supply design, there will be no space for computing units. The 800 VDC architecture is the solution to support the system's leap from kilowatt-level to megawatt-level.
In addition to the transition to 800V rack power supplies, this architecture also requires ultra-high power density and ultra-high efficiency in the voltage conversion from 800V to 1V. Only gallium nitride power devices (GaN) can meet these stringent requirements simultaneously.
To meet the power density requirements of 800 VDC, the power switching frequency must be increased to nearly 1MHz to reduce the size of magnetic components and capacitors. The typical switching frequency of current rack-mounted power supplies is up to about 300kHz. Increasing it to 1MHz can reduce the core size by about 50%.
Innoscience's Third-Generation Gallium Nitride Technology Has Decisive Advantages:
On the 800V input side, compared with silicon carbide (SiC), Innoscience's gallium nitride (GaN) can reduce drive losses by 80% and switching losses by 50% in each switching half-cycle, resulting in an overall power consumption reduction of 10%.
At the 54V output end, only 16 Innoscience gallium nitride devices are needed to achieve the same conduction losses as 32 silicon MOSFETs. This not only doubles the power density but also reduces drive losses by 90%.
Compared with silicon MOSFETs in the current rack architecture, using gallium nitride materials in the low-voltage power conversion stage of 800 VDC can reduce switching losses by 70% and increase power output by 40% in the same volume, significantly improving power density.
The low-voltage power stage based on gallium nitride can be expanded to support higher-power GPU models. Its dynamic response is improved, and the capacitance cost on the circuit board is reduced.
Power Integrations - The Only 1700V Gallium Nitride Supplier in the Industry
Roland Saint-Pierre, Vice President of Product Development at Power Integrations, said: "As the power demand of artificial intelligence continues to grow, adopting the 800VDC input solution can simplify rack design, improve space utilization efficiency, and reduce copper usage. As the rack power demand continues to rise, we believe that 1250V and 1700V PowiGaN devices are ideal choices for main power supplies and auxiliary power supplies, as they can meet the efficiency, reliability, and power density requirements of 800VDC data centers."
Power Integrations' InnoMux2-EP IC is a unique solution for the auxiliary power supply of 800VDC data centers. The 1700V PowiGaN switch integrated in the InnoMux-2 device supports a 1000VDC input voltage. Its SR ZVS operating mode can provide an efficiency of over 90.3% for the 12V system in the liquid-cooled, fanless 800VDC architecture.
Most commercial devices provided by manufacturers on the market usually have a rated withstand voltage of less than 200V, or their rated withstand voltage is between 600V and 650V. In the voltage range above 650V, only a few manufacturers have launched GaN HEMTs with a rated withstand voltage of 900V. It is difficult for commercial GaN HEMT technology based on silicon substrates to achieve voltage expansion above 900V because this requires an extremely thick buffer layer, which brings significant process challenges.
Therefore, applications that require wide-bandgap power devices with a rated withstand voltage of 1200V and above have been limited to using SiC switching devices. However, compared with SiC, GaN can achieve a higher switching frequency, providing a feasible path to meet the ever-growing power density requirements of applications such as AI data centers while maintaining high efficiency. The GaN HEMTs manufactured by Power Integrations using its proprietary PowiGaN technology have unique advantages and can achieve an extremely high rated withstand voltage (up to 1700V) in actual devices, making them a ready and attractive alternative to 1200V SiC devices and higher-voltage devices.
To fully leverage the advantages of GaN in 800VDC bus applications, a half-bridge structure with two 650V GaN devices connected in series is usually used, with a total of four 650V GaN devices. Although this stacked topology can operate at the high frequencies achievable by GaN, it brings several challenges, including increased control complexity, reliability risks due to input voltage imbalance, increased space occupation, and increased conduction losses, resulting in reduced efficiency and increased costs. In contrast, using a PowiGaN switch with a rated withstand voltage of 1250V in this application can not only significantly simplify the power converter topology but also fully utilize the characteristics of GaN - which is what makes it an ideal high-frequency power switch.
Using the 1250V PowiGaN cascode switch, power supply designers can be very confident that their designs can operate at a peak VDS of 1000V while meeting the 80% industry derating standard. For application scenarios where the peak operating VDS exceeds 1000V and reaches up to 1360V, using the 1700V PowiGaN cascode switch allows users to design equally efficient power supply solutions but at a higher voltage.
The above figure shows the schematic diagram of Power Integrations' cascode architecture. The 1250V/1700V GaN HEMT is a normally-on, depletion-mode device manufactured based on Power Integrations' proprietary PowiGaN technology. It is connected in series with a low-voltage silicon MOSFET to form a cascode structure to achieve effective normally-off operation, which is crucial for the safe operation of power electronics systems. Depletion-mode GaN devices are considered to have extremely high reliability because they do not require a p-type GaN gate layer. Therefore, they avoid threshold voltage drift and related instability problems, ensuring long-term stability.
PI compared the LLC topology with a fixed ratio of 800VDC and 12.5V output using 650V enhancement-mode GaN devices and 1250V PowiGaN devices. Due to the use of fast-switching GaN devices, both solutions can operate at a high frequency of over 500kHz with an 800VDC input. However, for the 650V stacked topology, several challenges will arise:
Input Voltage Imbalance: The input voltage imbalance during normal operation must be properly controlled. If there is an imbalance between the half-bridges, the stress voltage across the GaN devices may exceed the expected approximately 400V. Under this higher voltage stress, due to the current trapping effect in the 2DEG channel of the HEMT, the dynamic RDS(ON) degradation will become more obvious. These limitations highlight the reliability and efficiency risks associated with using the 650V enhancement-mode GaN stacked structure in an 800VDC input system.
Complex Drive Design: The stacked topology also increases design complexity, especially in the gate drive circuit. Each half-bridge requires a dedicated upper-tube driver and an isolated bias power supply, which further increases system cost, space occupation, and design burden.
Lower Efficiency and Higher Cost: When using GaN devices with the same RDS(ON), compared with the 1250V PowiGaN single-tube half-bridge topology, the stacked topology will generate higher conduction losses. This means that the 1250V PowiGaN design can use devices with an RDS(ON) value twice as high while still achieving the same overall efficiency and loss characteristics.
Additionally, compared with 1200V SiC MOS with an approximate RDS(ON), 1250V PowiGaN can achieve a higher-frequency LLC, thereby achieving a higher switching density.
Texas Instruments - The Most Comprehensive Product Portfolio
For the 800V power conversion architecture, Texas Instruments can provide products such as gallium nitride (GaN) power stages, digital power controllers, multiphase buck regulators, DC/DC point-of-load buck converters, hot-swap controllers, and isolated gate drivers, supporting efficient and high-density power conversion under the 800VDC architecture.
Currently, there are two conversion architectures for 800VDC. One is the three-stage conversion architecture (800V → 50V → 12.5V/6.25V → <1V): 800V is converted to 50V (efficiency 98%) through a 16:1 IBC, then to 12.5V (efficiency 98%) through a 4:1 IBC, and finally to the core voltage through a multiphase buck regulator (efficiency 92%), with an overall peak efficiency of approximately 88%.
The variable scheme is to replace the 4:1 IBC with an 8:1 IBC (50V → 6.25V, efficiency approximately 97.5%), and increase the efficiency of the multiphase regulator to 92.5%. The overall efficiency is still approximately 88%, and a lower input voltage can support a higher switching frequency, reducing the size, improving transient performance, and supporting backside mounting.
The other architecture is the two-stage conversion architecture (800V → 12.5V/6.25V → <1V):
64:1 IBC scheme: 800V is directly output as 12.5V (efficiency 97%) through a 64:1 IBC, paired with a multiphase regulator (efficiency 92%), with an overall efficiency of approximately 89%. This can eliminate the 4:1 IBC, saving size and cost.
128:1 IBC scheme: 800V is output as 6.25V (efficiency 96.5%) through a 128:1 IBC, and the efficiency of the multiphase regulator is 90%, with an overall efficiency of approximately 89%. However, there is a large current challenge (the current reaches 2.4kA - 3.2kA at 6.25V), requiring large-size conductors (such as busbars) and multi-module parallel connection to control circuit board losses.
Texas Instruments said that under the 800VDC architecture, the two-stage architecture has higher efficiency and greater power density, but the large current output of the medium and high conversion ratio IBC makes it difficult to control circuit board losses, requiring multi-module parallel connection.
Overall, power chips need higher and higher energy conversion efficiency (to reduce data center operating costs, heat loss, and air conditioning expenses), small size (the power components have limited space on the circuit board