Heavy desktop solutions have gained a new, fast-growing class: builds optimized for AI workloads, which structurally differ not only in having neural accelerators, but also in utilizing certain specialized components, including PSUs. We spoke with Karim Laskri, Business Director of FSP Group in the CIS region, about why new-type PSUs are needed.

High-performance AI PCs have burst into the class of high-performance PCs, which was previously represented by workstations and powerful gaming computers. The line between traditional workstations and high-performance AI PCs is becoming increasingly blurred and will eventually disappear—neural accelerators will become a mandatory element of all computers—but for now, such a division exists.

Interest in powerful AI PC variants has been growing rapidly lately, and under Russian conditions, at an accelerated pace. Data center issues are shifting user attention toward on-premise computing, while data leak concerns make the focus on on-prem particularly pronounced for systems built for training AI models.

Don’t you consider the emergence of AI PCs to be a purely marketing story?

The division of high-performance systems into traditional and AI PCs does have a marketing component—we are talking about the market, after all, and there’s no getting around marketing—but it also affects a number of technical parameters. Like other computing systems, AI PCs represent “mini data centers” whose design elements are created to ensure normal operating parameters for all components even under prolonged, specific high loads. We have previously discussed the importance of cases for high-load systems under such conditions; optimal placement of cooling system components—air or liquid—is an eternal topic at the intersection of technology and art. The main difference between the systems, of course, lies in balancing: a traditional workstation is optimized for versatile professional workloads—CAD, rendering, engineering calculations, video editing, etc., and in some cases, the balance is confirmed by a certificate from the software vendor. AI PCs, whose primary distinction is a neural processing unit (NPU), vary—there is specificity even within this class: there are solutions tailored for training AI models, and those oriented toward inference; modules for edge computing are even gaining popularity, as AI computations are increasingly moved to the network edge.

What components are needed for an AI PC?

A number of AI PC components—including power supplies—must match the computing workloads typical of AI operations. No matter how well balanced the system components are with one another, they need to be powered. Power supply units have undergone significant changes recently.

Do you mean increased wattage and higher energy efficiency?

Yes, AI computations require increased wattage—they are power-hungry. Energy efficiency is also becoming critical: the more energy a power supply dissipates during operation, the harder it is to keep temperatures inside a running PC case within normal limits. Let’s take our company’s new power supply lineup as an example. Its top model—the FSP Cannon Pro 3300—boasts massive wattage, which, as seen from the model index, is 3300 W. This is enough to power a high-performance system loaded with multiple graphics accelerators (for example, you can connect four RTX 5090 cards) under peak workloads. At the same time, headroom remains to power other components—from drives to cooling system fans. But there is one more requirement specific to AI PCs, which requires using models built on GaN transistors instead of traditional PSUs.

Why are GaN PSUs needed?

A key feature of NPU operation in real-world scenarios is rapid transitions from idle to peak power consumption, which take milliseconds. A power supply optimized for such solutions must manage to respond to changing conditions in time; otherwise, various unpleasant issues can arise, up to and including abnormal task termination. Therefore, AI PCs require PSUs with gallium nitride (GaN)-based transistors. The reason lies in physics: gallium nitride transistors provide faster switching thanks to high electron mobility and a number of other properties that are beyond the scope of this article. What matters to us right now is that such power supplies react to changes in external power draw noticeably faster than traditional silicon-based ones. GaN power supplies first appeared in 2018, but they were used in consumer chargers for mobile devices, and for their compactness rather than rapid adaptation to fast-changing loads. GaN PSUs for PCs debuted much later, just two years ago—right when AI PCs began to spread widely—but they still remain a niche product. The compactness provided by GaN is not a decisive factor for desktop PSUs, but it serves as a welcome bonus. The more free space inside the case, the better the airflow circulation and the more efficient the heat dissipation.

So it turns out the new PSUs are simply built on a new technology?

For GaN, as with any other technology, engineering packaging into a product is required. In doing so, a number of engineering challenges must be addressed. For instance, although GaN PSUs have higher efficiency, they still heat up during operation, and the reduced dimensions of such power supplies require additional design solutions for efficient and stable heat dissipation from a smaller case surface area. Issues related to minimizing electromagnetic interference from an operating PSU also had to be resolved. The high switching speed of GaN transistors leads to the generation of high-frequency electromagnetic radiation, which had to be shielded when designing the new PSUs to prevent interference with computing system components.

Do the GaN PSUs you offer differ only in wattage?

The principle is the same, but the models are different because we created them for different computing systems. We offer a line of such power supplies with common features typical of various solutions optimized for AI workloads—from consumer desktops to edge systems—though for obvious reasons, each subclass has slightly different requirements for GaN PSUs. Of course, there are differences in required PSU wattage. For example, the aforementioned flagship FSP Cannon Pro model is available in several wattage configurations—in addition to the maximum 3300 W, versions rated at 2500 W and 2000 W are available. But wattage is not the only difference among models in our lineup. There is the FSP Twins Pro server modification, which, as the name suggests, consists of two modules. The modules support hot-swapping, which is essential for systems requiring non-stop operation. The model is available in 1400, 900, 700, and 500 W versions. Hot-swapping is also featured in the FSP Twins CRPS, which are positioned for edge computing and therefore capable of operating in ambient temperatures from 0°C to 55°C and under reduced atmospheric pressure, enabling their use at altitudes up to 5,500 m above sea level. FSP Twins CRPS are available in configurations with a wide range of wattages—from 3200 to 1300 W. The FSP Flex Guru Pro targets consumer desktops—with a Standard Flex ATX form factor and a power range from 250 to 650 W, which is plenty for inference tasks. The FSP TFX 500 is similar, with a predictably TFX form factor, and we currently produce two variants: 500 W and 400 W.

Why do you believe the popularity of GaN PSUs will grow?

Russian enterprise customers need computing systems optimized for AI workloads, and in increasing numbers. Meeting this growing demand proves difficult: systems from key global manufacturers optimized for AI workloads are available to Russian enterprise clients with understandable restrictions. Russian vendors of white-box PCs are responding to the situation, though such products are more accurately called unbranded: building systems of this class in “garage assembly” conditions is problematic; furthermore, such computers require after-sales support. This leads to the growing popularity of local production of such solutions—small-scale or even custom one-off builds. Under these circumstances, there will be demand for specialized PSUs from a proven manufacturer, renowned for guaranteed quality and available on the Russian market in various configurations, complete with vendor service and established supply channels.