GaN server power platform helps drives the AI revolution

Navitas Semiconductor said that its CRPS185 3,200W 'Titanium Plus' server reference design surpasses the stringent 80Plus Titanium efficiency requirements and effectively satisfies the growing power demands of AI data centre power.

The rapid development and deployment of AI, including OpenAI's ChatGPT, Microsoft's Bing with AI, and Google's Bard, has infiltrated all aspects of people's lives. New power-hungry AI processors such as NVIDIA's DGX GH200' Grace Hopper' demand up to 1,600W each, driving power-per-rack specifications from 30-40kW to 100kW per cabinet. Meanwhile, with the global focus on energy conservation, emission reduction, and the latest European regulations, server power supplies must exceed the 80Plus 'Titanium' efficiency specification.

Navitas' reference designs dramatically speed up customer developments, minimise time-to-market, and set new industry benchmarks in energy efficiency, power density and system cost, allowed by GaNFast power ICs. These system platforms include complete design collateral with fully-tested hardware, embedded software, schematics, BOM, layout, simulation and hardware test results.

In this case, the CRPS form-factor specification was defined by the hyperscale Open Compute Project, including Facebook, Intel, Google, Microsoft, and Dell. The company's platform now delivers a full 3,200W of power in only 1U (40mm) x 73.5mm x 185mm (544cc), achieving 5.9W/cc, or almost 100W/in3 power density. This is a 40% size reduction vs the equivalent legacy silicon approach and readily exceeds the Titanium efficiency standard, reaching over 96.5% at 30% load and over 96% stretching from 20% to 60% load, creating a 'Titanium Plus' benchmark, vital for data centre operating models.

The platform employs the latest circuit designs, including an interleaved CCM totem-pole PFC with full-bridge LLC. The critical components are the company's new 650V GaNFast power ICs, with robust, high-speed integrated GaN drive to address the sensitivity and fragility problems associated with discrete GaN chips. Additionally, GaNFast power ICs deliver extremely low switching losses, with a transient-voltage capability of up to 800V, and other high-speed advantages, including low gate charge (Qg), output capacitance (COSS) and no reverse-recovery loss (Qrr). As high-speed switching decreases the size, weight and cost of passive components in a power supply, the company calculates that GaNFast power ICs save 5% of the LLC-stage system material cost per power supply in electricity over three years.

Compared to traditional 'Titanium' solutions, the platform design running at a typical 30% load can reduce electricity consumption by 757kWh and decrease carbon dioxide emissions by 755kg over three years. This reduction is equivalent to saving 303kg of coal. Not only does it help data centre clients achieve cost savings and efficiency improvements but contributes to the environmental goals of energy conservation and emission reduction.

As well as data centre servers, this solution can also be widely employed in applications such as switch/router power supplies, communications, and other computing applications.

"The popularity of AI applications like ChatGPT is just the beginning. As data centre rack power increases by 2x-3x, up to 100kW, delivering more power in a smaller space is key," said Charles Zha, VP and GM of Navitas China. "We invite power designers and system architects to partner with Navitas and discover how a complete roadmap of high efficiency, high power density designs can cost-effectively and sustainably accelerate their AI server upgrades."