LongCut logo

Project Deschutes Googles Open Spec CDU contribution to the industry

By Open Compute Project

Summary

Topics Covered

  • Seven Years of Liquid Cooling, Open Sourced
  • 2 Megawatts Per CDU for the AI Era
  • When Gemini Buys Your Data Center Hardware
  • Buy From Two, Qualify Ten: A Vendor Insurance Policy
  • Open Is the Key to Hardware Scale

Full Transcript

And good afternoon everybody and thank you for coming today. On behalf of my colleague Chi and um my many colleagues at Google, it is our privilege and

pleasure to be here with you today to present project dashes. This is Google's openspec CDU contribution to the industry

and we're very excited to be participating in OCP through the contribution of the project dashes specification.

Uh we're grateful to the OCP steering committee for their review and approval of the version 0.8 ate specification in

September of this year and as Cliff said it is now available through OCP for your review and feedback and she will share

more details about that in the call to action uh later on in the presentation.

The project dash to shoot specification includes the technical design details of Google's fifth generation uh CDU as well as the hot aisle

containment data center structure we call Redmond which is compatible with the CDUs and rear door heat exchangers to help enable the deployment of

highdensity liquid cooled IT racks. It

also contains the manufacturing quality, installation, service, and maintenance procedures that are informed by more

than seven years of scale liquid cooling experience at Google. And we're sharing these what we consider to be best practices to help enable fast

deployments of reliable equipment throughout the industry. And in addition to the technical details, the spec also

identifies the supply ecosystem for uh com system components that are provided by multiple vendors who are well known in the industry. Later on, she's going

to be sharing the supply chain strategy uh for competitive pricing, capacity, lead time reduction, and worldwide availability. It is our hope that the

availability. It is our hope that the industry will embrace project dissutes and we look forward to collaborating with and learning from the industry

through this spec release. So now let's take a technical deep dive into the project to shoot CDU. At a high level, the CDU is designed to be among the

highest performing CDUs in this form factor in the industry.

Uh from a thermal uh performance perspective, we are targeting to support 2 megawatt heat loads at an approach

temperature of 3° Celsius. And the the approach temperature is a common uh CDU performance metric which simply represents the temperature difference

between the two cold liquids that are available at the CDU. that is the cold liquid that is being supplied from the CDU to the IT equipment and the cold

liquid that is being supplied to the CDU from the facility.

Uh from the hydraulic performance perspective, we're targeting to support 500 gallons per minute at an available pressure head of roughly 80 to 90 PSI on

the secondary or the technology cooling loop side of the CDU. And these

performance targets are intended to support more efficient cooling of the next generation AI and ML equipment.

From a weight perspective, the CDU, a dry CDU would weigh about 5,300 lb or about 2400 kilos. And a wet CDU, that is

a CDU that is filled with liquid, would weigh about 6,900 lb, which is about

3,100 kilos. The CDU requires 380 to 416

3,100 kilos. The CDU requires 380 to 416 volts AC and would consume about 74 kW uh at full power. The heat exchanger is

liquid to liquid which means that liquid flows on both sides of the heat exchanger that is on the side that's connected to the IT equipment and likewise on the side of the heat

exchanger that is connected to the facility uh liquid loop.

Wetted materials are selected to be compatible with DI water as well as the water and propyline glycol mixture PG-25 which is commonly used throughout the

industry. And finally, the form factor

industry. And finally, the form factor is about 5 1/2 ft wide, close to 8 feet tall, and about 4 feet deep.

Now, we're very excited to see that uh there are already several liquid cooling industry players who have built project dash CDUs and actually have them here on

display at the OCP summit. And one thing you might notice is that while the images in this presentation are consistent with what is published in the OCP spec, they're not exactly the same

as what you may have seen out on the summit floor. And the reason for this is

summit floor. And the reason for this is that we've already been fast at work with the CDU vendors on implementing design changes that um will be reflected

in the next release of the project is specification. Nevertheless, the overall

specification. Nevertheless, the overall architecture and and system components are generally the same uh between what you're seeing on this presentation and

what is shown on the floor. Now with

that disclaimer, let's take a closer look at the design details. On the

secondary side of the CDU or the technology cooling loop side of the CDU, the architecture is such that the CDU contains two pumps, three heat

exchangers, four filters and pressure, temperature and other flow and other telemetry.

The pumps are driven by variable frequency drives and are intended to be redundant such that each pump can deliver the hydraulic capacity targets.

The heat exchangers are arranged in parallel and are optimized for thermal performance while minimizing pressure drop on both the primary or facility

side and the technology sides of the CDU.

mechanical fittings and uh flexible hoses are selected to take up the mechanical tolerances that are inherent in the brace plate heat exchanger

manufacturing processes uh and therefore enable the mechanical integration of the heat exchangers into the CDU assembly there on the primary side of the CDU

there's filtration as well as a valve which regulates the uh the flow that is consumed from the facility ility.

There's also a programmable logic controller or PLC for telemetry monitoring as well as to control the pumps to maintain a fixed differential pressure set point on the secondary side

of the CDU and to control the valve on the primary side of the CDU to maintain a fixed approach temperature set point.

But the controller can also support other controls methods. The metallic

piping is stainless steel and the CDU would be found in an enclosure uh that is not shown here.

Now, yes, the animation works. I was

kind of nervous about that. And so,

we've provided this this fun animation uh which shows how the city the CDU is integrated. And while the animation is

integrated. And while the animation is of a simplified assembly, it's it's shown here and it's intended to give you a flavor of the different uh subsystems

and components that are in the CDU. So

I'll just let this one finish up before I move on. So it's there just for fun.

And in addition to the CDU, the project to specification also includes the technical design details of the Redmond

hot isle containment data center structure.

Now this Redmond interface specification is provided to the OCP community to help enable the deployment of highdensity

liquid cooled racks. And as I mentioned early in the presentation, the structure is compatible with CDUs as well as rear

door heat exchangers behind each IT rack. And the rear door heat exchangers

rack. And the rear door heat exchangers would be anchored to the structure floor such that they would not require any physical or mechanical support from the

IT racks themselves. In addition to that, the structure supports the shoulderto-shoulder deployment of liquid cooled IT racks that are 28 in wide, up

to 50 in deep, and up to 90 and a half in tall.

And the design of the structure is such that racks of different depths could be deployed in a front justified manner where the containment plenums take up

the differences in the rack depth. And

similarly the uh racks of different heights can be deployed shoulderto-shoulder where the uh containment plenums take up the the

differences in height. The structure

also supports eight 600 amp bus ducts, four per side u that deliver power to the high power density racks in a four

makes three configuration and the structure height of 14 ft we found uh would be accommodated in many colo data

centers throughout the world. And with

that, I would like to turn it over to my colleague Chi, who will be walking you through the supply chain strategy of Project Dashes. Thank you very much.

Project Dashes. Thank you very much.

Thank you. Thank you. Hi.

Hello everyone.

Hello.

My name is Chi. I am the strategic sourcing manager in Google.

I see a lot of familiar faces down there. Thank you for coming to this

there. Thank you for coming to this talk. Hohi and I worked on project

talk. Hohi and I worked on project dissues generation one through four for many many years. We deployed total of

more than one gawatt of dissutes. So far

we achieved 59 99.999 reliability.

Now we are releasing dishes project five to the community.

My end goal is how do I green button and go back. Okay.

All right. Here it is. My end goal.

I want to build abundant capacity that provides worldwide availability within 10 to 12 weeks of

lead time at very reasonable and competitive cost.

Huhei started working on Dissutes 5 last year Q4 and we together released our version 0.75

to 10 suppliers earlier this year and then we released 0.8 specification to OCP community in

September.

I know since then my 10 supplier number is increasing and I can see this 10 is going to be increasing more and more. I

don't know how to count though but you may ask chi with that many suppliers. How do you manage them?

How do you define how do you decide which one you're going to buy your CDU from?

You know my answer, my ultimate answer will be I will go to supplier A's website, click

500, drop into my shopping cart. I pick

out my deep pocket and check it out. And

I will tell Gemini, Gemini, I actually totally need 2,300. Go figure which suppliers out of the 10 plus suppliers then you can buy from and come back to

me. Within five minutes, Gemini say,

me. Within five minutes, Gemini say, "Hey, all done. 2300." They will all show up within three months.

They will all arrive three months later.

No data on arrival.

No leaks, no clocks, no reliability ever problems. What a wonderful world we are going to live in.

But from here to the beautiful world we want to live in, I do have to walk very carefully and work smartly with all of

you.

So out of all the 10 plus suppliers, I'm going to categorize them into three categories.

I'm going to talk about the top and the bottom first and then come back to the middle.

So for Google internal deployments I am going to buy from AVL approved a vendor list.

Currently I have two spots and I want the lead time of these suppliers to be four weeks and out of one to three one

to two manufacturing sites three is better of course. Um, I work day in and day out with these AVLs. I know exactly

their whole supply chain and I know their manufacturing plan. I

know their control plan. I know their cycle time, their labor, their yield.

And we have a open book on the cost and we plan our capacity plan our capacity together

with our 18 to 24 months of forecast.

I use them for our own data center deployments.

So now let's go down to the category RVL recommended vendor list.

These are the suppliers. I really don't have enough resource time to manage them. However, I encourage everyone of

them. However, I encourage everyone of these RVLs, whoever wants to be our RVL's to make your samples to our

functional specification, test to our quality and reliability specification and certify yourselves with UL, EMC, and

any other related safety compliances that you need to comply to. Submit the

paper qual self-qualification data to me. A whole package. Hohi and the whole

me. A whole package. Hohi and the whole team and myself will be reviewing your paper call. If we feel everything is in

paper call. If we feel everything is in check, stamp it. You are our RVL.

We only have two AVLs. What if some quality excursion happens with these two AVLs? What if there is just some force

AVLs? What if there is just some force major event I cannot get enough or you know Google's demand is just exploded how many acts

within 12 months I just don't get enough capacity what do I do I rank all the RVL's

pick the best of the best of the RVLs and call them reserved AVL in the cases when AVL my true AVL

current AVL they're having issues I can quickly convert those reserved AVLs into AVL

within couple of I hope couple of weeks maybe a month or two I can start buying from the reserved AVL's

now question comes right chi we're here and you're not buying from us you're only buying from AVL and

RV reserved RVL then when can we have chance to sell it to you Google I know everybody wants to

ask my answer is yes you will be able to sell it to us when when you are ready when you are ready but t here's another

question if you are even not even buying from us we're not manufacturing How can we be ready?

So, well, this is the magic of the supply chain over here.

You can sell it to coloss with Google as the end customer. You can sell it to other CSPs. You can also sell it to

other CSPs. You can also sell it to other coloss with two minutes. Yeah.

sell coloss with other CSPs as end customers, right? The market you create,

customers, right? The market you create, you know our product, you know how to manufacture, you know how to certify, you created the product.

Let me speed up for components. I will only I make it

for components. I will only I make it simple.

AVL and RVL's. Okay, here I listed 16 major components. Currently we only have

major components. Currently we only have one supplier on our release the bomb. I

like to get it to two or three. What is

even more I really want to talk to this audience is all the CDU suppliers all the component suppliers don't reframe yourself join

this community qualify the more and more CDU suppliers and component suppliers expand the ecosystem make it more available. Remember, our

goal is to make it available within 10 to 12 weeks of lead time worldwide at a very competitive cost. Of

course, reasonable, right? Everybody is

here to make money. We want to make it reasonable. Call to action. We currently

reasonable. Call to action. We currently

only released 0.8 version.

there is room for you to input all the DFM great ideas how you can however make it more uh easier to manufacture make it

more efficient let us know we'll incorporate your information and we are targeting to release our 1.0 to zero specification in September, actually in

November this year. And thank you very much. We need a scale and open is the

much. We need a scale and open is the key to scale. Thank you very much.

Loading...

Loading video analysis...