CDFAM NYC 2023 · New York · 14–15 June 2023
Geometry Streaming: Hardware First Software for Digital Manufacturing
Abstract
Presentation Title: Geometry Streaming: Hardware First Software for Digital Manufacturing
Drawing upon his extensive experience at Metafold, he provides a comprehensive overview of the capabilities of geometry streaming, demonstrating its real-world applications and the benefits it brings to digital manufacturing.
This presentation is particularly suited for professionals, researchers, and academics in the fields of computational design, additive manufacturing, hardware design, and digital manufacturing.
Please note: The CDFAM Computational Design (+DfAM) Symposium is a distinguished platform for the dissemination of innovations and insights in computational design and additive manufacturing.
This recording is provided for educational and professional development purposes.
Transcript
From YouTube’s automatic captions, lightly cleaned; expect some errors. Each timestamp opens the video at that moment.
Read the full transcript · 2,680 words
0:05 I’m here to talk about geometry streaming, Hardware first software, on behalf of Metafold. I’m the CTO and co-founder, my name is Daniel. But before that, back to the to the 90s, you all know this story: it starts in the 70s when Blockbuster rose to video rental market dominance, culminating in the mid 90s with 9 000 stores globally and close to six billion dollars in revenue.
0:40 And then Netflix, in 97, read Hastings and Mark Randolph launched a mail order DVD rental business to compete with Blockbuster, and actually did very well. There’s often this idea that they weren’t doing well at the video rental market, actually, were doing quite well, but it, it was the release of their streaming platform in 2007 that I’m sure 100% of us all here today use that caused their meteoric rise to 232 million users worldwide as of this year, and huge profits as well.
1:22 And today the landscape is this: streaming is the de facto way to access video on demand. If you’re anything like me, you have an assortment of these different streaming services, and you flip between them all endlessly before going to bed. But the point is that the drivers for the video streaming market were things like fast internet, going from 500 megabits up to about a gigabit connection, 1000 megabits, mobile devices, and demand for content.
1:59 Interestingly, Netflix actually did a lot of its growth in the early, in the late 2000s and early 2010s, despite poor video quality, and I, I remember watching YouTube content at very poor resolutions and being very happy about it.
2:14 And so the proposition to you, and is also what we’re working on at Metafold, is that these are the same drivers, the same drivers that drove the video streaming industry also are driving the additive manufacturing industry: fast internet, mobile devices there’s so many folks that work from home now and maybe don’t have access to that massively powerful workstation, GPU devices are getting hard to acquire and expensive, expensive if you do, and there’s an increased demand for content.
2:54 And there’s one other driver that is made less explicit, and maybe less backed out, an economic fact, and that is convenience, and convenience is a very powerful force. I was just out of grad school when Netflix really rose to prominence, and as, as a student then, and also a software programmer, I was very adept at finding alternate sources to video content, like, was really good at it, and I switched over like almost immediately, right, like it’s just so convenient to access this from anywhere and quickly.
3:30 So we at Metafold really adopted this, this idea, and the nice thing is that convenience is a, is a notion that also motivates some really interesting technical goals. So these are the goals that we had when we were building Metafold: we want to run on any device, we want to get initial results from a query in about 200 milliseconds, so that’s about on par with what a streaming service does, it gets going after about 200 milliseconds products around our product, and the last one, we want to focus on dense meta materials.
4:04 Nothing to do with the video streaming service idea, you may wonder why we’re into that, and I’ll get there in a minute. The early kind of architecture sketches that we worked on in the early days of metaphor led us to this: we wanted to find a good way to deliver content in a way that was good for the human and also good for the machine, a kind of unified geometry representation, and so we built a GPU cloud compute service as a way to to hit all our target goals.
4:37 So this was a great idea, and that’s us now, we, we’re a cloud GPU compute service for the additive manufacturing industry.
4:48 So why meta materials, and this is a kind of a crucial part of our business. So I was at rapid last month, maybe a lot of you folks were as well, and I observed that that there were about 90, 90 of all the parts at rapid and form next before it were lattice based in metamaterials, and I actually did a count today, I kept a little counter, I think we’ve had about 10 talks today and nine of them had meta materials in them.
5:20 And so I would propose to you that these high frequency structures and dense meta materials are not a nice to have of the added manufacturing industry, but they’re in fact a defining trait, and, to be way more specific, if you’re not using microstructures and additive manufacturing, then you’re missing out on a big part of of the design space.
5:43 These are wonderful images, actually, from from HP, 3D Systems, BASF, again, I took these at rapid last month, just to hit this point home a little bit more. This is a paper that was out last week, and this is phenomenally interesting. The Gibson Ashby chart is something that all material scientists use, and it, it, you know, maps material properties to, on a chart, and this paper, one of the big findings is that the this this chart, and the properties that we associate with these lattice structures, break down after a certain scale, it’s just not accurate once you get to a certain length scale separation.
6:21 And so again, there’s a huge amount of potential and dense meta materials to build new products around, and that’s the sort of second big piece of what we do at Metafold. So we want to be convenient, we want to be fast, we want to run in all mobile devices, we also want to do this for dense meta materials, and this is challenging.
6:42 And so the next few slides are a little bit about the challenges that we hit in developing this tech. A quick calculation: so a pretty standard build size for for a lot of consumer, up to industrial printers, is about 200 mils in a number of directions, and let’s just say this is a DLP system, or something that works on voxels, at a 4K resolution, this gives you 36, almost 37 billion voxels.
7:15 And, okay, that’s a big number, but what do we do with this if you want to design with this? So maybe you want to represent this as a sign distance field, or or some other kind of you want to attach some kind of geometry primitive to this, this data size, you’re at least looking at something like 132 gigabytes in memory. So this is still well beyond your phone’s memory, your laptop’s memory, some workstations have this, not many, and it is certainly beyond anything that you would consider streaming.
7:47 So to solve this problem, programmers and developers come up with different representations. And so what’s a good representation for this kind of data, and this was a central technical problem that we, we’ve been working on. For those of you that aren’t CAD nerds, there’s a few very common ways to represent these kinds of shapes, they’re broken down into two parts: you have explicit representations and implicit representations.
8:15 So an explicit representation is something like triangles or polygons, this is the the the most well-known and widely adopted representation, for, you know, reasons unknown in 2023, but there it is. Also, in more standard CAD applications, you have spline surfaces, so nerves sort of embrace the fact that there are a lot of voxels and go with it.
8:41 You also have fancier representations, like sparse hierarchical grids and functional representations, which is also emerging as a really useful representation, and, like, you’ve done a lot of work on this, and others have done a huge amount of work on on F reps. The the point is that the whole dense thing, if you want to get at those material properties that have that length scale separation, a lot of these representations don’t work well, and you end up and this is a challenge that we had as well, and still have to some extent it’s very hard not to move the problem around.
9:25 So for instance, if you look at the picture on this slide, this is quite a dense representation, and you might think that a hierarchical grid or some kind of boundary volume approach might work really well here. The problem is there’s always more detail, there’s always a finer beam, or another thing that you need to subdivide into, and so the payoff from these structures deteriorates really really quickly.
9:53 Various file formats do improve matters, in the 3mf Consortium, Duann has done amazing work in the technical team, and the 3mf Consortium, really great work, to to improve matters on the triangle front, but there’s still no free lunch, like there the problem is still there, and this problem of compressing and reading dense data still exists.
10:18 So we opted for the streaming approach, so that we could not represent the whole structure in memory at once, but rather sort of drip it in as needed. And when this works, both for the modeling front, and also and this was a key driver from the beginning from Metafold also at the printer level, which most printers print by slices, that’s a great discretization of the shape right there.
Okay, so this following video is sort of like our acid test for the streaming performance, this was done not long ago, and I’m going to hit go, and the first result hits our target. So this is a just a normal browser, I am at my desk, I’m at a wire connections, about 500 megabits, servers in Virginia, I’m in Toronto, and a hit go, like this, start a lattice, and boom, so that’s streaming in at that level of detail.
11:17 And now the acid test here is to gradually increase the unit cell count down to the kind of length scale separation where we get some really interesting material properties. So from 20 millimeters down to 10, and you can see all the partial results updating as well, and from 5 down to two, compression struggling a little bit, rendering is getting crunchy, so we increase the the render resolution, things do bug down at this point.
11:50 So we’re now at a one millimeter unit cell, and about a 200 millimeter boundary box, but here we go, and it’s in just on the edge of passable, I think, from a geometry quality side, not quite, but that’s a, that’s about three million unit cells in the browser, streaming real time, and you can run this on your phone, there’s no compute on the device.
12:17 I was trying it here, the Wi-Fi is not great, cell is even worse, so at your desk with a wired connection, if you do want to try it. So that was our first big win, is the streaming architecture, and this is for the designer, this is for the user, but the same representation can be used to stream slices to the printer. So we expose a web API that lets you do high resolution slices, this is an 8K slice.
12:47 The second goal is to build out a geometry kernel that sits as a web service, that hits some of the really kind of gnarly CAD operations that you want to, that you want to do. And so in this example we’re going to start with the lattice in the same way and start to size it up. So a common, a common request is not to use, like, always to use the the circular beam sections, so we have a number of different braking, but we can size this up, and now we can start to do something really interesting.
13:22 So we have our approach lets us do something called redistensing on the fly. Redistensing is normally a really heavy compute problem, even with some of the really great algorithms, and so here I’ve set up a resistance operator, and the redistance operator is sort of the the gateway into some interesting CAD operations that are not easy to do.
13:46 And this, for instance, is a fillet. So this, you can take a lattice shape, apply a box profile, and you can fill it the whole thing automatically in the same framework. Filleting is one of the crucial CAD moves that removes stresses from the nodes and lets them flow around the node and not get sort of pinch points at these structures.
14:10 Another common pain point for a lot of implicit modeling is not being tied back to real units, so here’s a grading of this unit cell, and often when you grade the unit cell in an implicit modeler, you destroy real world measurements. So in this case we might expect the fillets to also distort, but because we’re using this redistensing operation, the fillets are exact.
14:39 So a bit of a deep dive into CAD nurdham, but for any mechanical engineers this will surely resonate with you. The other thing is that the everything I showed you is accessible through our API, and it works like a rest API, you have a token, and you make requests.
14:54 This request is just a Json file, I’ve exposed this in Python, but this is all that it is. You set up the volume, you set up some distance transforms, some some sdls, or design spaces, to sample, and you hit go, and you get results back, and you can build your own product with this, your own scalable web-based implicit modeler, and we have documentation.
15:22 And I couldn’t, you know, show off this at this conference, and for everyone, as a developer, this is truly astounding, more than the technical achievements is actually getting documentation together, yes, every call and response in the parameters, it is there for you to read.
15:47 With the API, though, we can do some really interesting things. So the the web app is kind of a sort of a one-to-one relationship between the user and the parts, but with the API we can do bulk queries, and we can process on a single GPU, on the web, we can process close to three to five thousand parts an hour.
16:09 And this meta ABC is actually a joint work with yushin Chen of U of T, built off the ABC data set. I think Daniella is around somewhere from New York, here, yes. This is a great data set of a lot of industrial CAD components, if you don’t use it, you should, and we were able to bulk process this data set, fill it with lattice elements, and you just create, save this out, and use it for training applications, algorithm robustness, and so on, so.
16:44 And we found with some of our enterprise clients that this is one of the things that they, they really want to do, and it sort of unlocks generative design for these, these parts, and internal products, or even commercially viable projects that they build.
16:57 However, all of this, I think, can be summarized at least for us, this is a recent win. We’ve been working with the Southwest Research Institute on cell scaffolds, cell scaffolds are something that our software really excels at designing, and when, are the researcher at Sperry came to us, the thing on the left was the prototype that they had.
17:16 And through our technology and our product, the the researcher could design new cell scaffolds in about four hours and print them, without exporting to STL, was a direct to printer streaming integration with the Sega, and out pops and industrially relevant cell scaffold with at least 100 times more surface area, and more surface area means more cells and a more industrial relevant part.
17:46 So for us, all this technology kind of accumulates into this one user success. So thank you, please try out our application.
More from Daniel Hambleton

Measuring Shape Fidelity in Generative CAD Models
CDFAM DC 2026

Accelerating Time to Market for Purpose-Built AM Software
CDFAM NYC 2024
Listen, search and cite
Listen to this talk as a podcast
More from CDFAM NYC 2023

Generative Design & Digital Manufacturing at NASA Goddard
Ryan McClelland · NASA Goddard Space Flight Center

Generate-to-Validate: Optimal Designs and the Need to Deliver Under Uncertainty
André Wilmes · Rafinex

Defining Performance not Just Geometry with Software Process + Material Science
Hardik Kabaria · Carbon

Reinventing the Wheel – Representing & Manufacturing Geometry with a Program
Mark Burhop · Siemens Technology

Interspecies Forms: Robotic Feedback Systems Mycelia Growth & Computational Form
Natalie Alima · Adidas / Biolab Studio

Designing Material Properties with Advanced Toolpath Generation
Alex Roschli · Oak Ridge National Laboratory

Design Automation Using Machine Learning & Computational Design
Rhushik Matroja · Cognitive Design Systems






















