Liquid-Cooled Power Shelf and 800 VDC platform for next generation datacenters

Okay.

Hello everyone. Um, so I'm Jared Huntington. I'm a power architect with

Huntington. I'm a power architect with Nvidia working on uh power basically from SST to chip and uh

Roy Moi CMO at Magme.

Yeah, I'm Chile and I'm here in Magnet and focus focus on the liquid cool power supply design.

Okay. Um so today we're talking about the um liquid cooled power shelves and our transition to 800 volt DC um and uh the work that uh Megmi is

doing with Nvidia on developing uh liquid cooled power shelves that um we can use for the uh e ecosystem.

So can't go too what's up. Yeah. So then uh hopefully

what's up. Yeah. So then uh hopefully everyone's seen seen a graph similar to this right we present this uh in several of our or several of our talks but this

is essentially what's driving the need for 800 volt um and it's essentially if we look at the power consumption over time right there's a a year-over-year

increase uh just due to the chips due to the GPUs um pulling more power for more performance but there's another aspect about GPUs that's unique

which is we really want to pack as many GPUs into one cabinet as possible because um we can essentially uh increase the performance if we keep all of those GPUs on a copper networking

domain. And so that's what's

domain. And so that's what's fundamentally different between GPUs and the classic web servers. So very quickly we can go from an NVL72 system and we

can double power to an NVL 144 system, right? And so that that doubling of

right? And so that that doubling of power can happen as soon as the packaging uh as soon as there's new developments in packaging. And so as engineers working on power, we don't

want to be the bottleneck in the system.

We don't want to be the gate for what's uh you know the next level of performance that we can get in the rack.

And so that's where we need to be prepared for power levels to increase very quickly. And so in order to support

very quickly. And so in order to support uh these compute racks, we've we've essentially uh we have several power generations. Our first generation with

generations. Our first generation with GB200, GB300 uh was capable of 145 kW.

Our second generation was uh basically capable of 330 kW and that's uh what we're using for Vera Rubin. And then now what we're working on for our generation

three of power platforms is uh 100% liquid cooled power shelves.

And with that, if we completely fill a rack and we push to the limits of a 50VT bus bar, we can support up to 570 kW.

And so, um, and then finally, we are working on native 800volt racks where we take 800 volt DC directly into the rack.

We're not prepared to show what that looks like quite yet, but, um, this generation 3 is, um, uh, is kind of a stepping stone to get us there. And so

from a a rack power architecture standpoint, what we're doing is we're taking in 800 volt DC and then we're converting it to 54vt DC on the 54vt bus

bar. And initially the the first step,

bar. And initially the the first step, the first phase in that development is a power rack where we take in uh AC and then we output 54volt DC or I'm sorry,

we output 800 volt and then we uh take the 800 volt and go to 54 volts. And so

what this allows us to do is longterm anything that needs to be 800 volt capable whether it's uh support racks uh

networking racks uh even our um our compute racks right we as long as it's been designed around the um MGX bus bar we can take it and we can make it land

in an AC data center or a um 800volt DC data center.

And so then uh for Gen 3, why are we moving to liquid cooling? So there's se several reasons why we're moving to uh liquid cooling for the the first is the compute just has to move to liquid

cooling. Um with a uh essentially with

cooling. Um with a uh essentially with all of the heat sinks that we need in the system, there's just not enough volume to fit everything. And the second

thing is once you uh once you actually take all of that um uh once you take the actual rack power, there's also just a a footprint

limitation on how much power you can dissipate in the rack. And so um even if we were to for our power racks start filling it with power shelves, there's still going to be a limit of you know

750 kilowatt to 1 megawatt of uh dissipation. And so then we need to move

dissipation. And so then we need to move to liquid cooling for that reason. The

next reason is all that fan power that we are dissipating, we can return that back to compute. Um, and also with uh power supplies specifically, we've been using a lot of electrolytic capacitors

for AC input current shaping. And so

putting those against a cold plate, right? It allows us to take down their

right? It allows us to take down their temperature, which um with the Aranius law, we basically every 10° C, we can cool those capacitors. We double the lifetime. So that's beneficial. And then

lifetime. So that's beneficial. And then

um we also can eliminate points of ingress. So anyone who's worked around

ingress. So anyone who's worked around power supplies and data centers, a common failure mode is dust shorts and debris and work that's happening. And if

we can completely seal the PSU, we eliminate those failure modes. And then

finally, what I'm showing here is the back of a a power rack just completely filled with power shelves. And as you can see, uh, trying to manage the the

hot air flowing out over all those power cables, it's just completely unmanageable unless we move to liquid cooled power shelves.

So then from a facility standpoint, uh, we we really want to improve the overall facility cooling efficiency. And so we can do that by eliminating chillers.

That's where we're moving to the 45° C liquid inlet temperature. Um and by moving to a dry chiller system with closed loops, we can basically um save

2.6 million gallons per megawatt per year when compared to um evaporative cooling, right? And so it's it's a much

cooling, right? And so it's it's a much more effective way to cool that allows you to go to a um or allows for a better

PUE. And uh we can also then with the

PUE. And uh we can also then with the better cooling capability increase the megawatt per uh meter squared. And uh

heat reuse becomes a viable option. And

I think one of the more interesting things is uh we can actually change the data center layout uh in new ways, right? We no longer have hotisle

right? We no longer have hotisle containment that you need to worry about. Um and you can now push racks

about. Um and you can now push racks right back to back or you can put uh even put racks into shipping containers more easily. So then uh up next uh Roya

more easily. So then uh up next uh Roya will talk more about um the work that Megat is doing.

Thanks.

So we all know about the benefits of liquid cooling coming into compute tray.

Uh but as we uh move forward to uh much higher and higher rack density, there is uh even more thermal challenges or concerns that needs to be addressed. Um

for us uh we solved this issue by making the uh power shelf liquid cooled. Um so

our 2U 110 kilowatt power shelf is 100% liquid cooled with zero air flow inside.

Um this is a complete fanless design.

So all that heat is extracted through the facility liquid cooling loop. Um and

so this fundamentally changes the whole infrastructure in the rack in the in the data hall. Um so there is no more uh

data hall. Um so there is no more uh traditional hot and cold aisle in the data hall. Um and relying on the

data hall. Um and relying on the high-speed fans and all the noise that comes with it. Those are all eliminated.

So coming into power shelf when we remove the fan and fan power and all the failure points that comes with the fan we actually enable a much higher density

power supply um and also more efficient use of the rack space and all the of course uh the white space that is there.

Um so when we combine uh electronics and uh electronics and liquid there's obviously there is a much more robust approach to the safety. So we added so

many more um liquid uh liquid uh sensors that is at power supply and power shelf level. Um so in case there is a leakage

level. Um so in case there is a leakage inside the power shelf uh it shuts down the power either power supply which is 55 kilowatt or 110 kilowatt which is at

power shelf level.

Um I want to uh say that there is more benefits to the cooling. So all those components like ECAP and semiconductors are always running at lower temperature.

So this increases the life of the components and also reliability.

Um and finally uh our experience with EV manufacturing, putting machines and everything that we have today enables us to introduce this technology much faster to the market.

So if you move both power compute and uh power shelf uh power converters both of them to liquid cooling uh our study shows that you can actually save on

capex up to 10%. And that's mainly coming from white space saving uh compact uh power conversion and also labor saving and all that infrastructure

that comes with the air cooling that is not needed anymore.

So we have a mature design that is at 54 volt DC. Uh we have released products

volt DC. Uh we have released products for MVDIMGX series. Um and also introduce all the power only racks which

is 800 volts DC. Uh our solution combines with uh all the DCDC power converters and modules that comes together with a compute tray. Uh as a

next generation our approach is to go to facility level. So that means with SSD

facility level. So that means with SSD that we are going to release next year uh we bring the whole uh 800 volt DC bus

bar to directly to the compute rack which is also liquid cooled. So that's

how we introduce and make it much faster to deploy um for any next generation of AI data centers.

Uh with that I'm passing to my colleague Chile.

Okay.

Okay. Uh I will introduce uh some um specific partial of design on the liquid

cool. So the first one is uh four uh 54

cool. So the first one is uh four uh 54 um 2 RU uh one uh 110 kilowatt liquid

cool uh ACDC power shelf. So this shelf uh is a two two RU and we have two PSU inside. So the PSU actually is a hot

inside. So the PSU actually is a hot swapable and so uh we have the shelf itself has dual AC um beeps and uh also

it has um bus clip for the DC output and and this shelf can support the media

uh MGX rack.

So for the liquid um for the liquid cool u specs that the uh the the liquid um

temperature is the range is from uh 17 degree to uh 50 uh 60 degree.

Uh the basic uh feature actually is uh uh the input is three-phase input and uh

also it has um 10k juice the internal energy for uh input current uh shaping.

Yeah, this is our liquid cool uh 800 volt DC power projects.

And for this rex actually it has eight um power shelves instead. So each power shelf can deliver 100 kilowatt power.

Um so the total power can reach to 800 kilowatt. Uh we have also the BBU um

kilowatt. Uh we have also the BBU um select you can use this uh BBU to uh

deliver more uh deliver uh 89 kilowatt discharge power and also it

extends the hold up time to 90 seconds.

This power shelf uh this power uh also have 300 W uh auxiliary power. We also

um provide the capacitor tray for this uh rack. So you can also use um

uh rack. So you can also use um capacitor tree to shaping the input current.

Uh this this page just show the upgrade from the air cooling to liquid cool liquid cooling. Uh you can see uh on the

liquid cooling. Uh you can see uh on the uh left side is uh air cool uh power supply. So uh it has a fan and it also

supply. So uh it has a fan and it also has um bulky aluminum hings. Uh on the right side is just an example of the

liquid cool power supply. So in the liquid cool power supply we removed the fan and and also removed the air channel. Um all the power components

channel. Um all the power components attached to the code plate at the bottom of the PSU. So with this configuration

um it can reach high power density and can get a smaller size. So with this benefit we can easily to put more

capacitor and to get higher energy storage uh instead.

Um since this um configuration removes the fan so we uh seal seal fan power and also we uh the uh power components

actually uh can um decrease the junction temperature such as the MOSFETs. So the total efficiency can be higher.

uh and also you know that the liquid um heat um transfer coefficiency is much

higher. So with this uh liquid cooling

higher. So with this uh liquid cooling the uh hot spot can be um suppressed so

it can realize a lower temperature and also high reliability.

Uh this just a compare between the EV code plate and the um IPSU code plate.

So the EV code plate actually is working and heavy and so we uh now we use the uh copper plate with embedded tubes. So

this um cold plate actually we can make it thin and also the copper uh tube is um uh how to see things is dis uh much

dissipation on the head. So we can use that one to get a lower temperature.

Uh this is what we are doing for the uh 100K 800 volt liquid cool ACDC power shelf. So we have two versions. One is

shelf. So we have two versions. One is

two RU with BBU version and another one is the F RU with BBU

and the two RU um without BBU actually is also um there is two PSU inside. So

the PSU itself is hot swappable. Um

the BB the F RU is just um uh that is just additional BBU modules uh to this

to RU uh sh power shelves.

Uh so uh the shelf itself have dual AC input beep and uh it can also support a

media um MTX uh rack and the liquid uh inlet temperature

range is um 17 C degree to 45 C degree.

uh the two um power shaft on the energy in internal energy is u 10 k juice.

Okay, that's uh this is another um product that is 95

kilowatt uh 40 uh 54 uh volt liquid cool DCDC per shelf. Actually this one is

used with um the 800 volt uh per shelf.

So this is the one uh shelf actually is one PSU. Uh so this one is also hot

one PSU. Uh so this one is also hot swappable.

Uh it can also support the uh media MTX rack.

So the input is 800 volt. So output is uh 54 volt.

Uh this is our um DCDC DCDC modules. Um

it's 800 volt to 50 volt. So the first one, the left one is the PDB module and it's liquid cooling and the total power

is um 30 kilowatt and the second one the red one is the 10 kilowatt DCTC module.

So I think that's all for my portion.

Yeah, if you have any uh question or if you in you are interested to our products, you can go to C19 booth and we can have a detailed discussion.

Yeah. Thank you.

Thank you.

Thank you.