Automotive virtual design reviews using VRED Quick Start on the AWS Cloud

DAVID RANDLE: Hi, everyone. My name is David Randle. I'm the global head of go-to-market for spatial computing at Amazon Web Services, or AWS. Today I'm going to be talking about automotive virtual design review on AWS, which features Autodesk VRED streamed to any XR device from AWS using NVIDIA CloudXR. So to talk about the use case first, I'd like to start by a video. And I'll narrate over this video since it doesn't have audio.

So within automotive design customers, design departments have been augmenting physical processes, such as 2D drawing, sketching, rendering, clay modeling, buck production, et cetera, with digital processes like digital sketching, rendering, 3D modeling, and visualization for more than 16 years in a quest to improve product quality, competitive differentiation, and product market fit.

These processes have been used in practice to increase stakeholder confidence during the decision making stages of product development, i.e., what does this cut line look like between the front and rear door of a car in order to restore balance if we move it by three centimeters, just as an example.

Now, unfortunately, over the past couple of years the COVID pandemic introduced a dynamic that created a bit of a problem for physical design reviews. And so a dramatic work dynamic shift that impacted the design review process happened because design reviews used to happen in person, not virtually. High-end graphics workstations would power real-time visualization of 3D versions of upcoming vehicle designs.

The stakeholders, including design managers, executives, designers, and marketers would congregate and review the vehicle designs presented on typically a large power wall or on high-end workstations-- sometimes even tethered with VR headsets. Now recently, within these past few years, VR-- or virtual reality-- and augmented reality devices, such as HTC VIVE headsets and Apple tablets, have been on the rise.

And these have been adopted within automotive enterprise quite widely, meaning there's now tens of thousands of mobile devices, for example, being deployed and managed within automotive OEMs. This is leading to new capabilities and augmenting the design process. Now, that proliferation of devices paired with the need to facilitate these design reviews amongst stakeholders who are located remotely has created the perfect accelerant for virtual design reviews, which we'll break down.

The video you saw was an example of how things might have historically been done. Notice that the participants are tethered to workstations. The workstations are located in a facility in this case. And what we're talking about today is how to power this through the cloud so that the devices can be wireless. And that you can get any members added to the experience without any specific hardware investment.

So the solution we've come up with to simplify this is a combination of Autodesk's VRED product, NVIDIA CloudXR, and our AWS EC2 G instances-- G4 and G5. To talk a little bit about the solution, we've packaged this in a very neat and tidy fashion to make it as easy as possible for IT staff supporting design groups within automotive customers to deploy and to also scale.

In order to do this, what we did was create something we call the Amazon Web Services Quick Start, which is a deployment guide and a CloudFormation template that sets up the utilities that are listed here, including CloudXR, Autodesk VRED Core, and NICE DCV so you can interact with the server. Now, in order to meet the desires of the automotive customers, these three components are required. So the first component, like I mentioned, is design review software, such as Autodesk VRED in this case.

And this has the capability to render a 3D vehicle in context at scale to multiple stakeholder points of view. And this is typically done from an on-premise server. But what we're talking about here is moving that workload to the cloud, and that's the VRED Core product that actually resides in the cloud in this case. The second component is a streaming protocol, such as NVIDIA's CloudXR, that compresses server-side rendering and decompress client-side images at low latency while transposing six depth of field-- or six degrees of freedom.

Sorry, not depth of field-- six degrees of freedom data back from devices, back to the server, and then to render the next frame from the adjusted point of view in real time. So that's happening at a 20 to 30 millisecond latency so that no user feels the difference. Wherever they look, it's sending that new position to the cloud. It's rendering that next frame and sending it back in almost an imperceivable amount of time. We call that motion-to-photon latency, by the way.

The last component is cloud or on-premise infrastructure to support this in a scalable fashion. And that typically leverages on-demand, real-time graphics capability with low-latency edge delivery. In this case, we're going to be talking about the AWS G5 instances or G-class instances, which are used to support this use case.

Automotive designers typically prepare the vehicles for evaluation in Autodesk VRED on desktop workstations and publish this content to VRED Core, which is residing on a Windows G4 or G5 instance on AWS. By the way, this can also be supported on premise through the AWS Outposts or closer to the location through our local zones. NVIDIA CloudXR, which is on the same instance, compresses that real-time rendered image from VRED and streams them to the NVIDIA CloudXR clients which reside on the VR headsets or tablets in this case.

Those are operated by the design review stakeholders from independent locations-- each from their own perspective-- as you saw in the video on the prior slide. And this happens simultaneously. What you enable is a virtual design review session to start when this happens by compiling these three ingredients. And what it's doing within automotive design studios is connecting stakeholders that otherwise would have to show up in person.

So it gives a tremendous amount of freedom at a great total cost of ownership-- which I'll talk about a little bit later-- to extend this capability to the design stakeholders. Now, to power this, I'll talk a little bit about the architecture because it's quite a simple one. But nonetheless, it's nice to share a little bit about what we've done with the Quick Start itself.

So like I mentioned, this Quick Start leverages Autodesk VRED Core running on an EC2 G5 instance. And it's using NVIDIA's CloudXR solution to stream the pixels being rendered in the cloud by VRED Core to eyeballs that are being-- eyeballs of the design stakeholders as they access it through virtual reality devices or mobile devices.

So what we have here is the primary user, let's say. This is the person that might set up or author the environment. And what they're interacting with is a G instance in a single-availability zone. It's got all the security protocols and public subnet set up. That's what the Quick Start helps with, by the way. And on that instance, there's Autodesk VRED Core and CloudXR, which is available through the AWS marketplace.

There's a CloudXR AMI. In fact, it was just recently updated to the latest version by this recording. So going to the marketplace to get that is an easy endeavor. And then we package this with a AWS NICE DCV so that you can actually visualize. You can see what you're interacting with server side to set up the environment here.

And once this is set up, Autodesk VRED is looking for an S3 bucket-- an S3 folder-- which is where the design stakeholders will upload their VRED scenes to. Once that's loaded into VRED, it does the rest. It's presenting that content, both for visualization purposes to the collaborators as well as collecting the right data from the right folders in the structure that we've created through the Quick Start.

And this is mirrored now. So we talked about a single user. I assume you're the first user setting this up, which would typically be IT staff supporting design studios. But this gets very interesting when you introduce collaborators. So collaborators, as they enter into the scene, there's an opportunity to list new collaborators through the interface within VRED. And each time a new collaborator is added, it stands up a new instance, either in the same availability zone or across multiple availability zones depending on the demands of the automotive OEM, in this case, for redundancy and security.

And so what you have now is an auto-scale mechanism that will accommodate each collaborator as they're added to this experience. And typically-- which we'll talk about in just a second-- these design reviews normally cater to, let's say, four to six stakeholders. Although, there's no real limit to this as long as the availability of on-demand instances is there, or the customer has reserved a set number of instances for them to use for this use case over the long term-- like, year plus.

So that's a little bit about the Quick Start. It is available on GitHub. So this is a no-cost asset that we're providing to the market. Again, it helps orchestrate the quick deployment of this solution within automotive design studios and other VRED users, even if they're outside of automotive design.

So to zoom in a little bit further, we're going to talk about the Amazon EC2 G5 instances, which are our latest class of graphics-optimized instances. There's five benefits I'd like to talk about here. The first is about the high-performance ability of the G5 instances, especially for graphics-intensive applications. What I mean by that is these instances deliver up to a 3% higher graphics performance and up to 40% better price performance than our previous generation G4dn instances.

They have more ray-tracing cores than any other GPU-based EC2 instance. They feature 24 gigabytes of memory per GPU. They support NVIDIA RTX technology, which is the real-time ray-tracing technology which VRED leverages. And this makes them ideal for rendering realistic scenes, running powerful virtual workstations, and supporting graphics-heavy applications. This is really the ideal instance type to support this use case. And we were very happy that this became available soon into the development of this Quick Start earlier this year.

So G5 high-performance graphics available on-demand at your fingertips. The other benefits include the fact that this instance type features a custom A10G, which is a Tensor Core-based GPU from NVIDIA built in partnership with us at AWS. They're the first in the cloud to feature this particular processor and, like I said before, deliver high-performance graphics for intense applications.

Each instance features up to eight of these GPUs-- so up to eight A10Gs depending on the instance type that you choose. And these come with 80 ray-tracing cores and up to 24 gigabytes of memory per GPU. They offer 320 third-generation tensor cores for AI/ML workloads, which obviously this doesn't tap into. But it talks about the flexibility and breadth of the G5 instances.

These come with the latest NVIDIA drivers optimized for Windows. So they offer the NVIDIA RTX Enterprise and gaming drivers depending on what type of workload you're using, at no additional cost. NVIDIA RTX Enterprise drivers can be used to provide high-quality, virtual workstations to a wide range of graphics-intensive workloads, just like the one we're talking about here with the Quick Start. The gaming drivers provide unparalleled graphics and compute support for game development as well.

G5 instances also support CUDA, cuDNN, NV Encode or NVENC, TensorRT, cuBLAS, OpenCL, DirectX 11 and 12, Vulkan, and OpenGL libraries. So very flexible, which is a dependency that we have on NVIDIA for providing the right graphics drivers for these instances. And that becomes part of the experience you get when instantiating a G5.

To talk a little bit about network and storage, G5 instances come with up to 100 gigabits of networking throughput enabling them to support the low-latency needs of machine learning inference in graphics-intensive applications. That 24 gigs of memory per GPU, along with the support of up to 7.6 terabytes of local NVMe SSD storage, enables local storage of large models and data sets for high performance as well. These G5 instances can also store large video files locally, resulting in increased graphics performance when streaming complex video files.

In this case, one can assume that the data sets coming from the design studios in the form of the VRED assets are well suited as well from a storage perspective based on the networking and storage capability of G5. Lastly, these instances are built on what we call the AWS Nitro system.

G5 instances are built on the Nitro system, which is a rich collection of building blocks that offloads many of the traditional virtualization functions to dedicated hardware and software to deliver high performance, high availability, and high security while also reducing visualization or virtualization overhead as well-- so another added benefit that a lot of people don't know about with the G5 instance.

Just to give some evidence that the people used using these instance types. For a very similar use case-- actually, almost the identical use case-- we have a customer called Varjo, who are one of the first to leverage the G5 instances. Varjo also integrate with Autodesk VRED and support the same use case through a much higher-end VR headset. So what we're providing here is a more accessible solution.

So this will work on any untethered headset, such as the Meta Oculus, Quest devices, or HTC Focus. Varjo provide their own very high-end VR headset and their own streaming protocol as well optimized for their headset. But nonetheless, they're powering this use case, so helping designers visualize vehicles using Autodesk VRED. And in this case, they say that for high-end VR/XR applications, such as the solution we're talking about here, the Amazon EC2 G5 instances are a game changer.

They're able to run professional applications in their signature human-eye resolution on their high-end headsets with three times the frame rate compared to our previous generation G4dn four instances, which they used before. And this provides their customers with never-before-seen experiences in quality when streaming from the server. So this is another point of validation that this is really the right instance type to power these types of use cases, especially the virtual design review use case.

Now, moving these workloads to the cloud has, obviously, another advantage beyond scalability. And that's what we call total cost of ownership, or TCO. The value of being able to leverage these instances on demand and at scale offers a pretty attractive value proposition to companies and our customers. To give an example of that, what we're seeing so far is that we're able to deliver this experience to automotive OEMs for less than the cost of a single VR class workstation per year.

And so if you think about the investment that's being made today to support designers with this type of hardware, by extending this capability you're not even reaching the point where you'd have to invest in yet another VR class workstation for that audience. So very attractive total cost of ownership here. The way we arrive at this, by the way, is if we assume that within an automotive OEM, you have a single design program.

That single design program is undergoing, let's say, three reviews per week-- maybe a bit more, maybe a bit less. You can do the math afterwards when I talk about the numbers. Those review sessions are about two hours per review, again, on average-- maybe sometimes less, maybe sometimes a bit more. And we can factor in about five concurrent participants.

At a cost of about $3 an hour for G5, times five participants, times two hours for each session, times about 156 sessions per year. You're looking at a total estimated annual cost of between $4,500 to about $6,000 to enable this experience for each design program. So like I said before, very attractive total cost of ownership-- typically less than the cost of a single VR class workstation.

And these numbers can be adjusted based on the $3 an hour published on-demand fee for G5. Obviously, with longer-term commitments for G5, that price can come down as well. And we would encourage any of the design studios to start to explore what this does in terms of adding new capability to their design programs.

I appreciate you being with me today and stepping through this talk about virtual design reviews powered by AWS, including Autodesk VRED and NVIDIA CloudXR. Again, my name is David Randle. You can find me on LinkedIn. And I look forward to talking to you next time. Thank you.