I wish I were going to attend the forthcoming IEEE Applied Power Electronics Conference and Exposition, which is scheduled to take place next week as I pen these words. There’s a buzz building around this year’s event like “I know not wot,” as it were.

APEC 2025 will be taking place March 16 to 20 at the Georgia World Congress Center in Atlanta, Georgia. When I say “Georgia,” I mean Georgia, the state in North America (one state over to the right from where I’m currently sitting in my office assuming I’m facing North), and not Georgia, the country that is itself sitting at the intersection of Eastern Europe and Western Asia. I really shouldn’t have to make mention of this, but people continue to get these two locations confused, two prime examples being ESPN in 2012 and the Trump campaign in 2024.

I used to be young and foolish (I’m no longer as young as I used to be). As a young sprout, for example, I remember thinking that only engineering was important, and I gave little credit to the necessity for marketing and sales. Now, I’m older and wiser, and I’ve grown to realize that engineering, marketing, and sales exist as interdependent members of a triad or troika. Like the three legs on a stool, it’s hard to maintain a comfortable balance if you are without one of your legs.

Much the same thing applied to the topic of power supplies. My interest has usually been focused on creating the digital logic devices that—at least to me—embody the “fun stuff.” Until recently, I rarely gave much thought to any components required to power my precious logic because that was always someone else’s problem. Now I feel like an old fool (but where are we going to find one at this time of the day?).

The thing is that you can create the most cunning XPU (CPU, GPU, NPU, TPU…) processing device that proffers the peak of performance, but it can provide this promised peak only if you can power it appropriately. 

Recently, I was chatting with Tim Phillips, who is co-founder and CEO at Empower Semiconductor. The folks at Empower have come up with a solution that will power today’s 1,000A XPUs, tomorrow’s 2,000A XPUs, all the way up to the 3,000A XPUs that—at the rate things are racing along—could be with us by the middle of next week (or, at least, much sooner than we might expect).

Consider an accelerator card boasting an XPU intended for artificial intelligence (AI) applications, remembering that a similar scenario applies to XPUs used to perform high-performance computing (HPC) tasks. A high-level view of the way things are today is as shown below:

XPU on accelerator card with traditional horizontal power delivery (Source: Empower)

The AI XPU (shown in blue above) sits in the middle of the top-side of the board. The remainder of the board’s top-side is occupied by power converters and inductors (indicated in brown). The above image is obviously not to scale. In a real-world implementation, the XPU can occupy less than 25% of the board’s real estate. Lest you think I’m exaggerating, take a look at a picture of an Nvidia H100 Tensor Core GPU mounted on an SXM5 module.

The H100 contains around 80 billion transistors. Nvidia’s next-generation B100 will contain 104 billion transistors, while the B200 (composed of two B100 dies with a 10 terabytes-per-second link enabling them to function as a single processor) will contain 208 billion transistors, and that’s not counting the transistors in the 12 high bandwidth memory (HBM) die stacks mounted in the same B200 package. Eeek! (And I mean that most sincerely).

The scenario shown above reflects what’s referred to as horizontal power delivery. The farther away the power converters are from their points-of-load, the higher the transmission losses. The worst thing we can do is to send current through a printed circuit board (PCB) laterally—it’s one of the lossiest things in the industry. Assume an XPU that burns 1,000W with a core voltage of 1V. That’s 1,000A right there. Now assume that anything up to 20% of the system’s energy can get lost just by being transmitted through the PCB. Further assume that power consumption is going to rise and core voltages are going to fall, which means current is going to go through the roof (in more ways than one). All this is rapidly becoming unsustainable from both thermal and electrical perspectives.

The farther away the power converters and components are from the XPU, the “bumpier” the voltage gets. If there’s one thing XPUs don’t like, it’s bumpy, ripply, noisy voltage supplies, which end up limiting their throughput. As a result, we need to slow things down and raise voltages to give us some headroom, which causes decreased throughput and additional power loss.

Why can’t we mount the power components on the bottom side of the board under the XPU, thereby providing vertical power delivery? Well, if you look at the image above, you can see a capacitor bank composed of thousands of small capacitors. Taken together, these basically act as a large, medium speed energy storage element. The fact that we need all these capacitors to be as local to the XPU as possible is one of the things that drives the power converters and inductors out to the edges of the board. In turn, the bumpy, ripply, noisy power caused by having the power converters and inductors out at the edges of the board helps drive the demand for the capacitors under the XPU. It’s like a Catch 22 situation (but not as much fun).

The other thing that would tend to prevent power converters being mounted on the bottom side of the board is their size. These accelerator cards aren’t used in isolation. There can be multiple accelerator cards mounted parallel to each other with only a few millimeters of space between them.

One of the reasons for all the inductors and capacitors mentioned above is that existing power converters cannot respond fast enough to changes in demand. If the converters could respond with lightning speed, the inductors and capacitors could decrease in size. Hold that thought…

All of which brings us to the chaps and chapesses at Empower, who have developed a solution that will bring tears of joy to your eyes. They start with the concept of FinFast Power Design, which involves placing power management into a processor-type silicon technology that employs special FinFET cells. Standard FinFET transistors can operate at gigahertz speeds but at low power. The guys and gals at Empower have evolved robust FinFETs that can handle high voltage, high current, and high power while still operating at high speed. When combined with innovative on-chip control logic, the resulting power converters can operate at very high frequency with low noise, responding to changing power demands 1,000X faster than conventional converter technologies.

But wait, there’s more, because the folks at Empower have also developed their own silicon capacitors and advanced high-frequency magnetic materials. Even better, they’ve developed custom power packages that are optimized for high frequency and power. The result is the Crescendo, which is only 6mm x 7mm in area and around 1.5mm thick (by comparison, a US quarter is 1.75mm thick). 

Crescendo (left) and multiple Crescendos mounted on the bottom-side of an accelerator card with a US quarter for comparison (right) (Source: Empower)

Do you recall the first image in this column? The one showing the XPU in the middle of the card surrounded by power converters and inductors, with thousands of capacitors mounted on the underside of the board. Well, now consider the same card powered by Crescendos as illustrated below,

XPU on accelerator card with Crescendo-empowered vertical power delivery (Source: Empower)

The lightning-fast power delivery response (we’re talking about nanosecond transient response) of the Crescendo minimizes the need for capacitors and inductors. Any capacitive and inductive elements that are used are created out of silicon and advanced materials that are presented in the Crescendo package itself. Furthermore, the Crescendo device is so thin that there’s room to add a laterally-mounted heatsink, thereby further increasing the Crescendo’s power delivery capability.

Tim tells me that the Crescendo platform offers “AI/HPC on-demand kW vertical power” with 3,000A+ of scalable peak power, 20x+ higher bandwidth, 5x+ higher power density, and up to 20% lower system power (which means we can be talking tens of megawatts of power reduction in the case of hyperscale data centers). 

Finally, feast your orbs on all that freed-up top-side PCB real estate (remember that the power converters and other components can easily consume more than 75% of a regular accelerator card’s available area). Now we have a choice—we could develop a much smaller card for high-performance edge applications, or we can create accelerator cards containing two, four, or more XPUs, although the thought of the currents involved in the latter scenario is making my eyes water.

If you are attending APEC 2025, the folks at Empower will be demonstrating their scalable on-demand true vertical power architecture for AI and HPC applications at booth #1445. Also, Tim will be giving a presentation, Vertical Power Delivery Enables Higher Performance from xPU and AI Products, on March 19, from 3:40 to 4:05 p.m. in room A412.

I only wish I could see all this with my own eyes, but someone has to keep the home fires burning. What about you? Do you have any thoughts you’d care to share on anything you’ve read here?