Aligned Data Centers (Aligned) today announced its new liquid cooling technology, DeltaFlow. The patent pending solution is designed to support the high-density compute requirements of next-generation applications and high-performance computing (HPC).
And it could have big ramifications as to how data centre capacity is booked for AI.
“We’ve been designing this for a while,” Phill Lawson-Shanks, Aligned’s chief technology and innovation officer, tells Capacity.
“We do a lot of work with chip manufacturers to understand how they’re developing their products, so we can evolve how we design, build and operate these building structures”.
Lawson-Shanks believes Aligned was ahead of the curve in terms of understanding how chip developments are changing data centre requirements.
The company acknowledges that legacy cooling infrastructure is not sufficient to meet future demands for HPC use cases, including artificial intelligence (AI), machine learning (ML), and supercomputers.
DeltaFlow allows Aligned’s customers to transition from air-cooled to liquid-cooled systems or deploy hybrid cooling systems combining both air and liquid in the same data hall.
It is specifically designed to streamline liquid-cooled deployments in a turnkey fashion by supporting customisation requirements and integration with current and future liquid-cooling technologies.
These include direct-to-chip, rear-door heat exchangers, immersion cooling, and whatever comes next, the company says.
DeltaFlow builds off Aligned’s DeltaCube product, a four-foot by four-foot cube that is capable of 146 kilowatts (KW) of heat rejection.
When stacked three high, it is capable of the same level of heat rejection as four computer room air conditioning (CRAC) units, in just four feet of space.
“If you look at legacy data centre design, you are able to facilitate 6.5 KW on average per cabinet and you can fit a megawatt (MW) in about 600 square metres,” Lawson-Shanks explains.
“But with the DeltaCube technology, we can fit a MW into just 186 square metres, with two rows of 10 cabinets at 50 KW a cabinet”.
DeltaFlow takes this even further, with the capability to cool up to 300 KW per rack.
Aligned says the combination of the two technologies means its customers looking to make the shift to liquid cooling (or integrate liquid-cooled systems in the same data hall as air-cooled infrastructure) can do so with ease and in a live environment.
This breakthrough has led Aligned to make the claim that customers looking to integrate liquid cooling into their existing deployments do not need to completely retrofit and retool existing data halls or construct new AI-dedicated facilities.
“We’ve always taken the view that chips are going to change, and density is going to change. We needed to be ahead of the curve,” he says.
Why are AI-specific data centres being developed?
“In traditional legacy data centres, the data hall is limited by the amount of cooling it can provide.”
Cooling technology already represents a significant amount of data centre opex, and even if legacy data centre providers wanted to install additional plumbing, Lawson-Shanks claims this is no easy task.
“Because of these limitations it's much more efficient for other data centre providers to build new buildings dedicated to accommodate AI workloads,” he says.
Aligned, however, claims that these requirements are met by its existing design philosophy.
Cooling AI infrastructure is a big challenge
Graphic processing units (GPUs) that are used for HPC are evolving rapidly.
“The designs are getting much more powerful all the time,” Lawson-Shanks says.
“A typical CPU by AMD or Intel takes about 250 watts of power. The latest Nvidia H100 chip takes 700 watts, so when you put 8 of those in a box, you radically change the heat profile of that server”.
Removing this kind of heat requires a different way of thinking.
“We designed the DeltaFlow specifically for this step change in cooling requirements and we designed it in such a way that it fits in the same envelope as the DeltaCube array,” he says.
This means that Aligned is able to install DeltaFlow in spaces that have already been assigned for additional DeltaCubes in its existing facilities.
It can also be attached to the same closed loop cooling ring already installed in its buildings, which goes out to airside chillers.
“It’s important to state that just removing the heat from the chip isn't enough because the rest of the server or the storage or the infrastructure, is still generating heat. You still need air cooling,” he says.
The DeltaFlow technology works with the DeltaCubes to address both issues.
“We have the DeltaCubes moving the cold air in and removing the heat. Now, we have the DeltaFlow specifically targeting the hottest part of the device, the chips themselves, so they work in harmony with one another”
Why is flexible cooling necessary?
“We're one of the very few out there that can cool AI infrastructure as it stands today, the way that Nvidia and others are designing their server technologies,” Lawson-Shanks states.
But as things are moving forward, he acknowledges that even Aligned needs to adapt.
In an AI or machine learning infrastructure, all of the servers are connected through InfiniBand, a high-speed fibre optic network, so they can all act as one unit.
But InfiniBand comes with the limitation that these servers can only be 50 metres apart.
“This means you always have to have a centre cabinet that just houses network infrastructure or InfiniBand switches,” he explains.
In addition to the liquid cooling required for the chips, the InfiniBand switches and other server equipment require air cooling.
Mixed rack densities within the same footprint require different levels and methods of cooling. Solving for the hottest or most powerful component is inefficient and likely unsustainable as more powerful GPUs are deployed.
“We can do that mixed density in the same physical space the way our technology works,” Lawson-Shanks says. “You need both air and liquid cooling technologies for this to work. We can do air across the board, but as the new chips are coming out, we know we've got to get ahead of that by providing specific liquid cooling technologies. That’s where DeltaFlow comes in”.
The reaction from the industry
“We don't talk about our clients and how they use their technology specifically,” Lawson-Shanks says. “But what I can say is our largest clients are very excited about this because what we have today can deliver what they need today”.
The Delta Cube array is capable of providing 50 KW a cabinet, which Lawson-Shanks says is sufficient in the vast majority of use cases today.
“But everyone's looking at what's coming. The way that chips are evolving we think it’s necessary to be able to provide 300 plus KW of heat reduction per cabinet. That’s what DeltaFlow does” he concludes.
