Even if a GPU in a data center only needs 700 watts to run a large language model, it might actually need 1,700 watts due to inefficiencies in the way electricity is supplied. This is the problem of Peng Zou and his team at the startup PowerGrille They are said to have solved this problem by miniaturizing and repackaging the high-voltage stabilizers.
The company claims that its new chiplets achieve up to 50 percent reduction in power consumption and doubling performance per watt by reducing the voltage conversion process and bringing it significantly closer to processors.
Contraction and movement of energy
Traditional systems provide electricity AI chips by converting AC power from the grid to DC power, which is then converted back to low-voltage (about one volt) DC used by the GPU. With this voltage drop, the current must increase to conserve power.
This exchange occurs close to the processor, but in a low voltage state the current still travels a significant distance. High current flowing over any distance is bad news because the system loses power as heat proportional to the square of the current. “The closer you get to the processor, the less distance the high current has to travel, and thus we can reduce power loss,” says Han Phuc Lewho explores power electronics at the University of California, San Diego and has no affiliation with PowerLattice.
Considering the constantly growing power consumption of artificial intelligence data centers“It's almost become an overwhelming problem today,” says PowerLattice's Zou.
Zou thinks he and his colleagues have found a way to avoid this huge loss of power. Instead of relieving the voltage a few centimeters away from the processor, they figured out how to do it a few millimeters away, inside the processor case. PowerLattice has developed tiny power delivery chipsets that shrink in size inductorsvoltage control circuits and software-programmable logic in a chip about twice the size of a pencil eraser. chiplets are located under the processor chassis substrate to which they are connected.
One of the challenges PowerLattice faced was how to reduce the size of the inductors without changing their capabilities. Inductors temporarily store energy and then release it smoothly, helping regulators maintain a stable power output. Their physical size directly affects how much energy they can use, so making them smaller weakens their effect.
The startup solved this problem by creating inductors from a special magnetic alloy, which “allows us to operate the inductor very efficiently at high frequency,” Zou says. “We can operate at a frequency a hundred times higher than a traditional solution.” At higher operating frequencies, circuits can use an inductor with much lower inductance, meaning that the component itself can be made from less physical material. The alloy is unique because it retains better magnetic properties than comparable materials at these high frequencies.
According to Zou, the resulting chips occupy less than 1/20 the area of modern voltage stabilizers. And each is just 100 micrometers thick, which is roughly the thickness of a strand of hair. This miniature size allows chiplets to be located as close to the processor as possible, and the space savings provide valuable space for other components.
PowerLattice chipsets will sit on the underside of the GPU case and provide power from the bottom.PowerGrille
Even though it's small in size, the patented technology is “highly customizable and scalable,” Zou says. Customers can use multiple chips for more or less patching if their architecture does not require it. “This is one of the key differences” between PowerLattice and voltage regulation problem, according to Zou.
According to the company, using chiplets can reduce an operator's power requirements by 50 percent, effectively doubling productivity. But this figure seems ambitious to Le. He says 50 percent power savings “may be achievable, but this means PowerLattice must have direct control over the load, including the processor.” The only way he thinks this is realistic is if the company has the ability manage power supply in real time depending on the processor workload – a method called dynamic voltage and frequency scaling– which is not in PowerLattice.
Facing competition
PowerLattice is currently conducting reliability testing and validation before releasing its first product to customers in approximately two years. But bringing chiplets to market will not be easy, since PowerLattice has well-known competitors. Intelfor example, develops Fully integrated voltage regulatora device partially dedicated to solving the same problem.
Zou doesn't see Intel as competition because, aside from products that differ in their approaches to solving the energy problem, he doesn't believe Intel will provide its technology to competitors. “In terms of market position, we are a little different,” says Zou.
A decade ago, Le said, PowerLattice would have had no chance of success because the companies that sold the processors only guaranteed the reliability of their chips if customers bought them. power supplies Also. “For example, Qualcomm could sell their processor chip, and the vast majority of their customers would also have to buy their own processors. Qualcomm power management chip, because otherwise they would say, “We don't guarantee the entire system will work reliably.”
However, now there may be hope. “There's a trend of what we call chiplet adoption, so it's heterogeneous integration” says Le. Customers mix and match components from different companies to achieve the best system optimization,” he says.
And while established suppliers such as Intel and Qualcomm may continue to have an advantage over established customers, smaller companies (mostly startups) building processors and artificial intelligence infrastructure will also need energy. These groups will have to find a power source, and that's where PowerLattice and similar companies can help, Le says. “That's how the market works. We have a startup working with a startup doing something that actually competes and even rivals some of the larger companies.”
Articles from your site
Related articles on the Internet
