0:00 Ignore that. Hey guys, I’m Ryan O’Hara. I am at a company called Alloy Enterprises and I’m our VP of business development. Fancy title, I’m wearing a suit, but you know, it’s great to be back here and amongst my engineering brethren and maybe even some former colleagues. I was in the air force for 20 years as a developmental engineer working a lot of advanced technologies for the air force composits high temperature materials and was able to stand up one of the first metal additive manufacturing labs for the air force.

0:36 Spent some time at NTOP and a previous contract manufacturer and additive manufacturing and I’m excited to be at alloy as as the idea of manufacturing evolves and also how we tie that to computational design. So, thanks for having me and happy to answer any questions after the talk. Okay, so at all we’re here to talk and and I got a bit of a unique talk that’s a little bit different.

1:01 We’re showing some amazing things in computational design and and one of the things that I’ve been in my career is that always been in the intersection of like what the realm of what’s possible and then what we can actually make. And I’m excited to talk to you a little bit about how we’re leveraging computational design. But there’s been some great presentations and and I don’t want to be mean to anybody some complicated ways of doing things.

1:23 And I’m going to talk about computational design from a real practical and simple standpoint and as it meets reality, right? And so one of the things that we’re focused on here is to deliver advanced cooling solutions enabling unmatched performance efficiency for basically cooling, right? And the idea here is that we’re at the intersection of manufacturing and design. We’re Okay, got it. Doing a whole rhythm down here.

1:55 So, we were founded in 2020. We’re a 50 person team. About 75% of us are engineers and we’re based in Burlington, Massachusetts, just outside of Boston. You’re more than welcome to come by if you’d ever like to to see the things in reality, but you’ll see them here today, too. And we basically own end to-end design, manufacturer and delivery of parts for customers that with thermal management requirements.

So like everybody here or a lot of people here, you guys are all smart engineers and and doing great things. Well, sometimes people just have problems and they don’t know how to solve them and that’s kind of where they come to they come to us, especially on the thermal management side. And I think there’s some tools that we’ll talk about today that you guys are using every day that are causing some thermal management headaches for people in the industry.

2:44 So specifically, I’m going to talk about thermal management with liquid cooling. And it’s going to be singlephase and it’s only going to be with a fluid. There’s a bunch of ways to deal with thermal management, but we’ll talk about why we’re focused on liquid cooling. And the big one here is this is a Nvidia Blackwell chip B200. And it and for some folks this is making my life easier in terms of emails and doing my homework and writing essays, maybe even some talk tracks for presentations, right?

3:12 But this is a Nvidia B200 chip, right? And it is powering the AI revolution that we’re seeing today. Trouble is it gets hot, right? And if you’re a chip designer in Nvidia, you know, I’ll share some personal thoughts. I think maybe the chip industry didn’t really think about all the implications of putting a lot of transistors in a very tight space. Right? So this chip is about a 50 mm by 50 mm package, right?

3:39 And it can produce quite a lot of heat and we’re talking about thousands of watts, right? So what we need to do to mitigate that heat is we need to cool it. And it turns out that liquid or air cooling isn’t good enough anymore. And so now you have to use a product like what we manufacture. This is a coal plate. And that cold plate sits the top of the GPUs, right?

3:58 And it turns out so those get hot, that gets hot, and then we’re seeing even some new designs at some new chip conferences where there’s eight of these in a single rack, right? And then you’ll notice that one rack is part of a a whole bunch of racks in here and it’s all getting really hot, right? And so and we talk about how hot they’re getting is, we need to talk about this increasing demand, right?

4:22 So you’re seeing like a 30% increase in compounded annual growth in terms of thermal design power or what we call TDP, right? And so the chip manufacturers are trying to respond to this and I would almost say that they’re kind of late to the game a little bit, right? And so they’re trying to produce parts that are or they’re, you know, how do I deal with this thermal problem and everybody’s kind of scrambling to figure it out.

4:45 And then more importantly, there’s a whole ecosystem, you know, not just Jeet and Nvidia, right? That there’s the people that make the motherboards, there’s the people that make the racks, there’s the hyperscalers. All these people are having to deal with thermal management at different different levels, right? And the key here is remember we talked about that 50 mm x 50 mm package. So, you know, today’s market 1200 watts, they’re talking about nextg chips up into the 3600 watt category, right?

5:10 So, like I’m old, you know, I finished college in 2000. So, I still had incandescent light bulbs, right? So there’s a 100 watt light bulb I think of those and that’s like 50 of them, you know, all kind of like in one spot in this little tiny area, right? And just one light bulb could be pretty warm, right? And if you’re were a kid in elementary school, you had a little chicken coupe and a you know, one light bulb kept all the chicken the chicks hatching and growing, right?

5:34 So now figure 50 of those in that small space and then you multiply that by like eight of those on a motherboard and then a whole bunch in a rack, it becomes a big problem, right? So one of the things that we do we provide value in a in a couple different ways and the key is you know we talked about that end toend ownership and we have basically a whole library of geometry right that we can put into a part right through the design then we can iterate on those parts that would be our computational design piece where we have some folks here you can see we have enthology here we also use a number of other tools to to to design and those things are actively iterating and And then ultimately we have fabrication right so when you can do the design you can iterate and then you can fabricate so that your process is optimized that becomes an ideal situation and I’ll talk about some of our process and how we own that on the manufacturing side.

6:29 So a couple key things though is that for our data centers and from like a direct liquid cooling standpoint we basically can provide thermal management with low pressure drop. And if you’re not familiar with pressure drop, that is the when you’re passing fluid through a passageway, it kind of wants to slow down, right? And that slow down and that flow ends up in a lower pressure. And so if you have a pump, that means you have to have a bigger pump the more pressure drop you have.

6:56 So if you can mitigate that, that’s going to be an ideal situation, especially when you’re in a stra space constrained environment like a data center, right? And also in that regards, you know, bigger pumps or smaller pumps in this case would actually save you a bunch of power, right? And then you start stacking up all these GPUs. That becomes an issue. The other piece here is like electronics and water generally don’t mix, right?

7:21 So we don’t want those things to leak. So that’s a big deal. And you’d be surprised how many of these liquid cold plates actually do leak. The other key here is we do cooling not just in data centers but in other industries like semiconductor tests, high energy lasers etc. And oftentimes you we have to re operate in either a high pressure environment or a vacuum environment. And then the other piece here is that GPUs are just the start.

7:48 So it turns out the GPU gets hot and then well if you’re going to have a GPU with all this throughput now the memory modules are getting hot and then all the other peripherals it just becomes a cascading problem of things right and so alloy is like kind of we’ll say lucky or well positioned that we have a capability that can help across the space there so alloy started in 2020 and the original idea was that we were kind of born out of the u u added manufacturing space in Boston, right?

8:19 And a number of our founders were were at companies like Form Labs and other folks. And so, the idea here was though is that additive manufacturing specifically with lasers is often very expensive and especially in metals, right? And the whole idea was how can we make a low pro or lowcost scalable process that could actually deliver large quantities of parts, right? And the idea here is that you’ll see in number one there we have a service prep but the key here is that we have a coil of metal in this case aluminum 6061 right and we also have a copper 110 so two materials and the whole idea is after we surface prep that we go through a construct process that’s basically we take that coil and we laser cut it we also can apply an inhibition layer with a basically a binder jet where we can prevent parts from sticking together and ultimately we’re going to stack those parts parts together and we’re going to put them in the third part of the process which is called diffusion bonding.

9:13 We’re going to produce that part by adding in a controlled environment heat and pressure to form a solid part. So much like limestone, you know, you get sediment layers on the limestone over time that heat and pressure forms, you know, a limestone or rock, right? In this case, we’re forming bonded aluminum. And then what’s unique about that is that there’s no filler materials. So, in cold plates, previously, binder, excuse me, brazing is used where you basically machine a part and then you put a solder or some sort of, different material together and basically put in an oven, vacuum brazing oven, and it forms a solid part where that filler material bonds to both sides.

9:54 Well, we’re doing this as a tonic layer, so there’s no different materials in there. And often in brazed parts, they they’re prone to leaking, especially over thermal lots of thermal cycles. Okay. And then finally we just do a heat treat for that part. For aluminum and copper is a little bit different. Okay. So manufacturing in the real world. So we have real parts. That’s our that looks like that’s our surface prep, right?

10:16 And then this is our construct machine. And we have the bonding machines. So we have five production lines on the the cutting side and the bonding side. So we have plenty of throughput to do thousands of parts that you would need for a data center. Okay. And we have a whole bunch of so I’ll get on to the fun stuff. That’s all like boring kinds of things here.

10:35 But the key here is that you know there’s a bunch of talks here about AI leveraging for for compute. And then you know some of the things that go in here is that we have heat basically and coming from all of these different parts inside of the data center right or in the server rack. We have it on the GPUs. We mentioned previously the DIMs. Now we’re talking about interconnects for PCIe cards and all sorts of other stuff here.

10:57 Solid state drives is another big one. Things that I didn’t even knew know. Get hot, actually get hot and it’s a problem, right? Especially when they’re in that space constrained environment. Okay, great. My video is playing. So, that was a big question mark. So, another key here is once you take that pressure drop and, heat mitigation, you talk about thermal resistance and, and now we have lower pressure drop.

11:21 And the key here is that we’re taking kind of this network or a vascular type structure. So, this is a CT scan of one of our parts. It’s actually this aluminum part that I have here. Happy to manu talk about that more. And then these are actually some sheets associated with the part. So it’s just sheet metal, right? And we have multiple layers with different types of geometry that we bring together and that’s where the diffusion bonding process happens to form the solid one.

11:46 But what we can do with this layered approach is we can have different types of geometry of different sizes and that’s where our computational design comes in. Okay. So this is optimized for different length scales. And the other key, we’re going to talk about two areas of computational design. We’re going to talk about how we massively paralyze the flow and then how we actually get it to high surface area geometry which we call capillaries.

12:08 So that’s kind of this next slide, this aluminum part that I showed when we kind of take a multi-tered approach. If you look at this this no laser on this here, but then the center image there, we basically have a series of passages that form like a network together to bring flow in, right? And then the key is you’d be surprised as I learned especially in the defense and industrial space where I focus on we get a lot of parts that need thermal management but they’re basically it’s called gun drilling basically have a bunch of drilled holes and it’s pretty crude way of manufacturing but it’s very cheap but can often provide a lot of turbulence and and different types of things at those intersections.

12:43 So one of the things that we can do with by owning the design manufacturing process is sometimes just making flow smooth or your passageways smooth, right? Put a nice bend, put a nice flow. And that’s going to reduce your pressure drop a lot because your flow is going to be nice and laminer, right? And we’ll talk about laminer flow here in a little bit. The other key here is that we have micro channels.

13:03 And you’ll see that so at the top we’re bringing flow in and then as we get down, we’re getting smaller and smaller to these that vascular structure where we’re getting these nice micro channels, right? And then there we go. So couple things that we have are microchs or micro capillaries, right? And the idea here is that with our process, we just have laser cutting sheet metal. So if you’ve had any parts made like or really any kind of, you know, Duann’s second grill out after this that’s all sheet metal, it’s laser cut.

13:35 That’s a pretty common thing, right? And so what we can do is we can actually just do single line curve cuts in our sheet metal to provide a microch. And depending on the material, for like aluminum, it’s 180 microns is our min channel width. And for copper it’s about 50 microns. Right? So we can make these ultra fine passages where the flow can come through. But often times when you have flow especially in a narrow channel it’s going to want to slow down.

13:59 Right? So if I had a long micro capillary that’s going to cause a lot of pressure drop right in this scenario we actually just want nice thin and short channels right and by bringing in that network of flow or massively paralyzing that and you’re going to produce something that looks a lot like this. So we basically are bringing the flow in and instead of having long microchs that might be 25, 50, 100 millm long, we basically have micro channels that are about 5 mm.

14:23 And when we do that, we get a nice laminer flow over that part, right? And then and but we still get a lot of surface area. Okay. And then apparently I’m talking faster or slower than I thought. So the other thing here is that from a computational design, we showed a lot of fancy things. But often times all I need to know is where’s the heat and I’m going to put as many micro channels as I possibly can in that space, right?

14:48 That would be like the first iteration and then I might depending on where my inas and go outas are I’m going to tailor those microchs to maybe avoid hot spots or provide a little bit more density where maybe I’m not getting as much flow. Sometimes on the ins and outs we don’t get to control that and right and so you’re kind of have to deal with the cars that you left and that’s where our real computational design is.

15:11 It comes from and we we do some things there and that’s kind of where things like nTop come in where you can use spatially varying fields to generate or to from simulation and you can modify the geometry there and then you can use other tools like like seaman’s flow EFD we do that for the CFD side and we can even tie those things into like neural nets and other types of things to tailor that microch geometry I think the key here is just for reference to when you talk about micro capillaries depending on the manufacturing process and this is key where we own it is that you can get different size microchs with our process.

15:46 We can make the as small as possible and then by paralyzing it we get a lot of benefit from that. And this is just a good example of what we’re doing on the dim cooling side. And I think that’s it for me. So appreciate everybody’s time. I’ll be around all day and tomorrow.