To maintain warfighting overmatch, coordinate deployed forces, and enable new warfighting capabilities, the DoD is actively looking to new programs such as Joint All-Domain Command and Control (JADC2) strategy to ensure warfighters have maximum situational awareness. This is driving the development of networks linking the cloud, command posts, combat platforms, and dismounted warfighter – including the addition of vast numbers of sensors – backed by big data processing, artificial intelligence, and machine learning.
With adversaries contesting all domains, including the electromagnetic spectrum, tactical network programs must forward-deploy technology to ensure continuous availability of critical information resources. Advances in GPU-accelerated sensor data analytics software, combined with SWaP-optimized computing, address the most pressing of these challenges, making it possible to deliver the benefits of modern C2 architectures to the network edge.
Join Nima Negahban, CEO and Co-Founder, Kinetica and Charlie Kawasaki, CTO, Curtiss-Wright PacStar, in this educational webinar exploring how high-performance analytics software and modern computing can be deployed at the edge.
Transcript
Bob Ackerman
Good day, everyone. I'm Bob Ackerman, the Editor-in-Chief of the Signal Magazine and I'd like to welcome you to our Signal Media webinar titled Intelligent Sensor Data Analytics at the Tactical Edge. In today's webinar which is sponsored by PacStar and Kinetica, two experts will discuss how forward-deployed technology can bring vital networking services to the warfighter. Now, those of you who have participated in past Signal webinars are familiar with our audience interface. Throughout the presentation, webinar attendees are encouraged to submit questions electronically through the ask a question box on the webinar console. When our experts are finished with their presentations they will answer as many questions as time permits during this hour-long session. Also, the resources tab offers resources to attendees throughout the event.
Today's first presenter is Charlie Kawasaki, CTO for PacStar, a Curtiss-Wright company. In his 16 years with the company, Charlie has been part of a team that won tactical networking equipment awards for several defense department programs across the services. All told, he brings more than 40 years experience to the table in a number of networking, cyber security, and systems integration disciplines.
Joining Charlie as presenter is Nima Neghaban, CEO and co-founder of Kinecta. Having developed the software and the core vision for the Kinetica platform, Nima also leads the company's technical strategy in roadmap development while managing the engineering team. He has extensive experience in developing big data systems across a broad spectrum of market sectors.
So with that, I'd like to turn the presentation over to Charlie Kawasaki. Charlie?
Charlie Kawasaki
Thanks, Bob.
In order to better coordinate deployed forces and enable new capabilities the US Army, Air Force, Navy and SOF are actively looking to new programs such as JADC2 and ABMS to ensure warfighters have maximum situational understanding and high-speed decision support. These programs will adopt a variety of compute and bandwidth-intensive technologies, increasing the use of sensor data, video, big data analytics, artificial intelligence, and machine learning to deliver the command and control information that warfighters need. The software enabling these capabilities is increasingly developed to run in the cloud which itself might reside in a range of data centers ranging from the large commercial services to the DoD's regional hub nodes. These are located in separate strategic regions and used by deployed Marine Corps and Army units to access information from theater tactical networks around the world. Next slide, please.
One tricky challenge around this is that tactical organizations deploy in distributed architectures with remote teams frequently served with poor wide-area connectivity and with each node or team of a network distributed over large geographical areas. This diagram from US Army illustrates this well. Each node or team is potentially disconnected from communications with intermittent service and frequently limited bandwidth. This is a situation that requires unique capabilities when contrasted with typical enterprise and commercial communications networks. In particular, the remote nodes or teams need local compute and storage resources to provide necessary information technology support even when disconnected from the WAN. This means advanced IT-based capabilities must function in a distributed and disconnected manner. Next slide, please.
One example of advanced capabilities the DoD is looking to field is the cyber SU or Situational Understanding Program run by US Army PM Mission Command. They are developing an integrated system intended to give commanders a full view of the battlespace in the cyber domain, delivered in phases. First with visibility into our own networks and systems. Then including a more full view of the cyber domain including EW and threat intelligence. Then lastly with integrated response actions, decision support, and orchestration with a full view of multi-domain operations as you can see in this roadmap. Next slide, please.
In this more detailed view of cyber SU, you can see the program intends to fuse NET OPS information, cyber threat, EW, CEMA, weather, and open-source data such as social media to develop a full picture. In the picture on the right, you can see the intended end user and in-field commander with a common operational picture. Next slide, please.
Here's another example program - the US Marine Corps armored reconnaissance vehicle. Circled in red, you can see a full complement of onboard sensors and sensor processing systems including with integration from UAVs for ISR and EW support. Next slide, please.
This slide shows a US Army call for white papers for demonstration systems supporting dismounted soldiers supported by AI fusing disparate data streams into common operational pictures, in turn, controlling robotic systems. Next slide, please.
That call is consistent with the US Army vision called the adaptive squad architecture integrating multiple sensing software and analytical technologies from multiple DoD programs to equip soldiers with the most advanced information-based warfighting overmatch possible. Next slide, please.
One of the most visible efforts towards this vision is PMIVAS, shown here, implementing augmented reality headsets backed with extensive information resources in an expeditionary context. The technology that goes into making these decisions available comprises not only the head-mounted displays but also includes wireless networking, advanced analytics, sensor fusion, data storage, and computing and it all needs to run in a tactical setting. PacStar has focused on creating the necessary compute and network infrastructure to enable these applications and we've collaborated with Kinetica to provide the IoT and sensor database and processing engines to make sense of massive quantities of data in the field.
I'm going to turn this presentation over to Nima next to introduce the Kinetica technology. Then I'll pick it back up and show you how we make it available for tactical use. Nima, over to you.
Nima Neghaban
Thanks, Charlie.
Okay, I think it's really instructive to understand the value of Kinetica by going back to our origin story. So in 2010, we were working on a really innovative DoD research program which had the goal of consuming about 200 different real-time data feeds and then delivering a common query capability to analysts and developers, so they could rapidly create, you know the best analytics, the best applications and give those to soldiers in theater. You know, at the time Hadoop was all the rage, now SQL is all the rage and you know. They adopted a common solution pattern that we still see today which was:
- we're going to make a query inventory,
- then we're going to create, you know elaborate data structures to try to service that query inventory and
- then scale-out that data structure as our data capacity needs grows.
What we saw there was the classic set of problems that we still see today which was very, very fragile system where that query inventory was not nearly enough for the analysts and developer needs so as future requests came in their ability to service those requests was extremely limited. Hardware fan out and expense went through the roof. And finally and most fatally that the latency of the indexes themselves, of those data structures themselves, just fell behind far too much to be of any value to the warfighter. So, we were there and in a former life I had a lot of experience with the GPU and we had an idea which was, you know for the past 50 years databases have been designed with computing as a scarce resource. Right? Now with the GPU you know, you have this device where compute is an abundant resource.
What if we design a database where we flip that equation, right, and we're trying to now leverage compute as an abundant resource and instead of trying to make these elaborate data structures we optimize for being able to consume in real-time, do complex queries simultaneously and leverage this compute and so it was with that idea that we were able to become the spatial and temporal fusion engine and ultimately overlap engine for this program. And that, you know, allowed us to get seed funding and really take off as a software product company.
And I think it's just informative to just show this is actually an application built on top of Kinetica from 2011 and you'll see here like the analyst has drawn, you know, a little triangle and what you're gonna see is over 200 different feeds being correlated through space and time and then being visualized in near real-time for the analysts in their session. So, I'll just hit play here and you get a sense of what I'm talking about. So, Kinetica is unique as a database because not only does it do high-speed data processing and query calculation, it also does visualization which for situational awareness, and with anything that has a geospatial context is very important. So now, you see here this analyst has immediately gotten a sense of all the entity tracks that ever flew, that ever passed through that little corner street that they outlined. And similarly, this is the next step of that where we're powering a heat map so they can understand concentrations of activity. And again any data set that has an X and a Y, Kinetica can power this type of analysis and visualization and it can do it across many data sets that are disparate and have no kind of prior planning or correlation.
So here you saw, the analyst drew a bounding box, and then they're saying across about 200 different tables that are increasing simultaneously, I want to understand how activity is tracking over time and show me any concentration areas, right. And so, not only are we doing all the actual data processing, we're actually doing the visualization here.
So moving on, you know, I think the fed space you know, really has been far ahead of the commercial space as it relates to being able to really understand the challenges of sensor data and really understanding what is the true value that can be derived. Right? The program that we worked on 10 years ago is what I'm seeing commercial enterprises starting to think about now, right. Which is, you've got this new evolution of data where you've gone from a limited set of highly structured, highly pre-correlated data, and transactions to semi-structured, loosely correlated things like interactions to disparate highly structured, highly geospatially, and temporally encoded observation data that's being generated by a wide variety of sensors. And really what we're talking about here is any type of sensor reading a timestamp, and an X and a Y. So with Kinetica, you can get a much deeper understanding of your entity or your situational awareness across many, many disparate data sets through just time and space as well as our ability to do correlation through other intermediate data sets.
And taking this further, like I said, fed has been really ahead here, ahead of the commercial space. In the commercial space for a long time, it was really just about connectedness of sensor information, getting a red light or a green light, and getting the latest version of that sensor reading. Where fed has been and continues to strive for is going beyond that. Being able to fuse sensor data to get a deeper insight to be able to take those readings and really do aggregations and analysis over time to give a much richer, deeper understanding. And when you can couple that vision with a product like Kinetica, you're able to deliver much more powerful services, data services for your enterprise or your government organization especially around situational awareness. Our ability to do the processing, the visualization, as well as power the real-time decisioning really is a game-changer for federal organizations.
So what is the hard part of this? You know there are tons of OLAP engines out there. There are tons of databases out there. Aren't you just talking about data integration and doing simple aggregates, following aggregates after that? No, what we're talking about is highly compute-intensive joint operations that most databases either don't support like these as-of-temporal joins or are not able to deliver the performance profile that's necessary for them to be of use for the modern enterprise. So we're talking about geo joins, we're talking about these temporal as-of joins, and we're talking about n-way equijoins where you have many, many tables in between that are going to be able to correlate to sensors on the left and the right. This is not your primary key/foreign key join, this is something that takes a lot of compute power and that's what Kinetica really excels at.
And if you take a look across the landscape today, some of these solutions may sound familiar to you, because this is how people have been trying to solve this. They've been trying to solve this with streaming platforms like Kafka and k SQL. The problem there is you have very limited or no historical context ability in your query capability. You have limited time series analysis and usually no geospatial capability. You have your big legacy and modern warehouse players, but again they really have been built to rely on these elaborate data structures and require that data engineering, that pre-planning that is time-intensive, is organizationally intensive and really that's the only way that they can kind of service this type of processing pipeline. On top of that they have issues just keeping up with the rate of consumption of data. Forget the query, just being able to consume the sensor feeds fast enough is of issue for a lot of your kind of legacy and other warehouse players and then many of them just have very limited geospatial capability, limited-time series capability. And then you have kind of what we were doing 10 years ago in the government which was trying to put together a very complex point solution that was hand-rolled and leveraged several left layers of data infrastructure. And ultimately with this, it's the same thing we experienced 10 years ago. It's a lot of build, it's very expensive and time-consuming from an op-ex and development perspective. If things go wrong, it's difficult to troubleshoot because you may not have the same team as the original ones that developed it and if you want to improve your capability, it becomes increasingly onerous for that team because it's usually something where they've not been set up to just build something and treat it as a product. It's been built as a point solution and that's really kind of this dichotomy that's always problematic for the long-term future in that enterprise.
So you know, if you have to just label what are the key things that you're not getting today? You're not being able to make your decisions fast enough. You're not able to really analyze with context. And you're not able to fuse - you're not able to get a deeper correlated picture of the entity that you're trying to understand. If that vanity might be a vehicle, it might be a person, it might be a device - whatever the entity is - you likely have many other disparate data feeds that you can correlate through at least space and time. And if you were able to do that easily and on the fly, you'd have a much richer understanding and you'd be able to do better decisions.
So if you had to say what does this technology look like, what should it look like, that's going to be able to solve these issues for me? It should be able to consume at speed. We know one of the greatest challenges of this problem is just keeping up with the rate of data being generated. I need to be able to consume at speed across streaming platforms like Kafka and my data lake and data warehouses, like your snowflakes, S3s, Azures, and bob stores. And then I need to be able to do data engineering and data fusion on the fly, very flexibly. I should be able to join disparate data sets that are large and continuously growing at a click of a button or at the speed of thought. I shouldn't have to get together with 2 or 3 or 4 different teams and try to make some sort of key convention so that when it's time to try to fuse data I have some key that's easily matchable. If I can power this in a way that's as I'm thinking, it could really unlock a lot of insight for me and my enterprise. And then finally, you know, having more than just your standard OLAP operators in SQL but really expanding it out to space and time where you have a rich set of geospatial operators and you have a rich set of temporal operators. Just time alone can reveal quite a bit about the entity you're looking at by looking at: okay here are all my sensor feeds, let me see within one-minute buckets or one-second buckets how activity is linking up across the different sensor feeds that touch that ending. And that's what Kinetica is.
Kinetica is the database for space and time. We specialize in consuming data at speed and scale, using that data easily and in an agile way so that the developer analyst can really do that at the speed of thought, and then giving rich and powerful OLAP geospatial and temporal operators that they can use to not only analyze and get situational awareness but develop decisioning pipelines such that they can feed events up their enterprise stack.
How are we doing this? How are we powering this? So what's unique about Kinetica is that 10 years ago we made a decision to not port MySQL and not port PostgreSQL. We said you know we're going to build this from the ground up to take advantage of this modern hardware and so that meant basically maturing our kernels and our kernel algorithms to be optimized for what's called vectorization. So this isn't your kind of classic MPP task-level parallelism where you're just you're doing the same kind of fundamental algorithms, just doing, you know, you're just splitting it up in an embarrassingly parallel way. This is fundamentally rethinking the algorithms and the intermediate structures that allow you to do you know those things like a join or a group by such that they are fully optimized to leverage some of the vectorization engines like the GPU and now more recently Intel devices with AVX.
As we've been going through that journey, the origin story that I talked about, the next major event for us was we won an IDC award for our work in the DoD and USPS just came knocking on our door because they saw a press release around that award. They said you know we just put a breadcrumb-emitting device on every mail carrier, on every truck, and we have no way of querying this. You know we spent millions and millions of dollars creating this vast sensor feed - we have no way of understanding it because our current data-processing platforms simply can't service the need. They can't consume it fast enough, they can't query it in ways that are interesting to us. So we were able to deliver in about a 6-month time frame a production-grade, real-time situational awareness and fleet management fabric that sits across all of USPS. It's got about 30,000 users daily and it's tracking all mail carriers, all mail trucks, all pieces of mail, and giving facility managers across the nation situational awareness into what they can be expecting, how their fleet is doing that day and doing that in in a real-time fashion.
More recently at FAA, we had a team, so there's the FAA data challenge. This was basically across the FAA, the mission being to show us your best ideas. A team from General Dynamics leveraged Kinetica to create a drone pattern of life decision engine, a situational awareness engine, very similar to you know what I showed you previously. What we've created out of the DoD space - that ability to take different disparate data sets and then do real rich pattern of life analysis in real-time and there's a small screenshot to the right. That is quite a different capability than what you're seeing out there in space today, where the windows are much more limited, your ability to correlate is much more limited. Here, we're really giving everything to you at your fingertips. We're actually keeping the granular readings. We're not doing roll-ups or anything where we lose the actual data fidelity and so it really unlocks a tremendous amount of power for an analyst. Simply with standard BI tools and I'll show you an example of that with cyber data.
At NORAD, we're powering probably one of the largest IoT platforms in the world. All of the radar sensors across North America are being fed into Kinetica and there's ML models and analysts and developers then querying Kinetica to power decisioning and insight and analysis. If a bird or a plane is flying in North America, rows are getting created in Kinetica so you know hundreds of billions of rows a day being created.
And before I switch the demo for our conversation today at the forward-deployed site, what's truly important is being able to power a lot of capability, a lot of you know analysis and compute power in a small footprint in an appropriate power profile, and with our ability to leverage the GPU and vectorize processing, that's really something that's unique to Kinetica. We started from the ground up to build for these devices, and with the GPU especially, you're able to deliver a huge amount of compute capability in a very small power footprint and physical footprint.
So with that, I'm gonna switch to my demo real quick. Hopefully, everyone can see my screen. What you're seeing here is just our basic admin tool and you can see that this is a four-node cluster of Kinetica, so Kinetica is a distributed system. I'd just like to show folks the data. I'm going to do a cyber demo. I'm just looking for the NetFlow table. We have our schema here and then we have our NetFlow table, so this is 18 billion records. If I refresh again, it's being ingested into live. It'll probably refresh in a second here once the router sends more data. This is across the Kinetica network so it's constantly being streamed into. If you look at the table itself, you'll see it's a standard table so relational constructs you know, you make a table, you set a schema, you begin ingesting at speed and scale. It should be very natural for your developer and analyst teams. So what I'm going to do from here is switch to the lightweight BI tool that we bundle. With this NetFlow table, we're going to be showing off this analysis with a standard BI tool, like the one we bundl,e but this could be Tableau or this could be custom developed. We're showing you a heat map of activity, we're showing you top talkers, we're showing you categorization of events, and we're also generating a graph based on source and destination IP. I'll zoom in here. You can do full pattern of life here and you know really use Kinetica as a very very capable threat detection and hunt engine for the cyberspace just off even a single NetFlow table, but of course, with that, you can use our join engine, like we've been talking about, and link Netflow to application logs and any other sensor feeds that you think are relevant, open source. Whatever you have access to, in that moment, is available at your fingertips.
So one thing I like to just show off real quick is you can see this is all of the traffic across the globe. You can see the distribution of data events. As packets flow through our routers, we have rules probably like everyone does, firewall rules. When things are denied, if they're deemed malicious by the router, it's the NetFlow record is tagged with a denied value and you can see the distribution across the globe. Denied is a small fraction of the overall traffic but what i'm going to do now is I'm going to just filter through all of my Netflow traffic time. I'm going to look at China and so what we're doing is we're doing all that OLAP, all that geospatial filtering, all that graph traversal and you can see the distribution now is quite different. You can see the denied is actually the overwhelming majority, and I can see here are the top talkers in that region. Let's see the top talkers that are just being tagged with denied. This is classic pattern of life analysis but we're just doing it at a much much bigger scale. We're doing it with geospatial being a first-class citizen so that you can really use that as a correlative and analysis pivot through your workflow.
From there just to kind of show other data sets, this is a taxi feed. This status is not real-time. This is a geospatially rich set of asset tracks from New York City Taxi. Here again, I can do very rich analysis. I can start with geospatial, you'll see the rest of the OLAP pattern of life update. I can go and start with a different entry point into my pattern of life and at the end of the day this is all SQL, so you can use standard tools you can use Tableau, Power BI, if your organization has any BI platform where it is leveraging SQL, you can leverage Kinetica in a lift and shift manner.
I'm going to switch it back to Charlie.
Charlie Kawasaki
Great. Thanks, Nima.
As I mentioned in the introduction, PacStar leads the market in tactical networking and computing systems with a large installed base in ground networking and increasingly in vehicle-managed solutions, be it ground air and naval. So we're ideally positioned to deliver the kinds of technologies you just saw where they're needed at the tactical edge. Next slide, please.
Here's an example of one of these types of programs. It's the US Army network cross-functional teams (CPSV) which is currently developing prototypes of its command post vehicles. Essentially, mobile data centers that in future iterations could support local cloud. These are medium tactical vehicles or trucks that carry a small data center's worth of servers enabling data resiliency. Next slide, please.
As I implied before, as forces develop greater dependence on cloud-based services denial of wide-area access to the cloud due to electronic warfare becomes a critical problem. A key to delivering advanced tactical capabilities and ensuring their continued availability is to vastly increase the capability of networking and compute at the tactical edge, replicating critical data and services in mobile cloud infrastructure. As such the DoD is looking to deploy cloud replication between remote computing nodes and upper echelons, as shown here. Next slide, please.
In another example, the Navy recently issued an RFP to support tactical cloud analysis including an approach for providing remotely deployed cloud or processing services for sensor enrichment in case the tactical unit becomes temporarily disconnected from the tactical network. Next slide, please.
So how does PacStar, now a part of Curtiss-Wright Defense Solutions Division, make situational understanding technology available to the tactical edge? We make three major types of products that apply.
- First on the left is our small form factor rugged hardware, popular in many US and partner tactical network programs.
- Second is IQ-Core Software, down the middle, a single pane of glass to manage tactical networking equipment. This is widely deployed particularly across large numbers of programs at US Army PEO C3T. It helps the DoD field complex systems while reducing the setup time, management time, and troubleshooting challenges in systems based on enterprise and cots technology.
- Lastly, on the right are our PacStar Integrated Solutions combining hardware. IQ-Core Software and an extensive suite of third-party technologies into complete deployable solutions. Next slide, please.
Our integrated tactical solutions start with our small form factor hardware family that has a large number of module types including routing and switching modules with embedded Cisco technology including our new top-of-the-line Cisco-based 10 Gig switch. Next build, please. It also includes a broad array of server types based on Intel Xeon processors with up to 120 gigabytes of RAM, 16 CPU cores each, and up to 122 terabytes of storage on our storage server. We have NVIDIA GPU-enabled servers for AI and video processing applications and we'll dive deeper into those in just a few slides and you'll see where Kinetica comes into play. Next build, please. We also have radio gateways for radio over IP, radio relay, radio crossbanding, and voice management, and a wi-fi AP for local and meshing wireless communications. Next build, please. We also have a variety of adapter sleds for Type 1 encrypters, hardware security modules, and tactical radios - handy when we can't physically modify the devices and need to integrate them into our system. Next slide, please.
Using our family of hardware components, we assemble the solutions into a wide variety of systems and use cases from small backpack systems to large command post, vehicle-mounted, and rack-mount configurations. These are rugged, backed by multiple UPS options and extensively MIL-STD tested. We typically work with partners - next build, please - who integrate the systems onto the platform and with various peripherals to create complete program solutions. Next slide, please.
To create complete solutions, we first start with PacStar 453 and PacStar 454. The PacStar 453 is a five-pound, high-performance server based on Intel Xeon D processors with up to 16 cores, 128 gigs of RAM, and 30 terabytes of storage. Importantly for Kinetica, this server includes an NVIDIA T1000 GPU with 896 CUDA cores providing 2.6 teraFLOPS of computing. It requires only 110 watts and is ideal for highly SWaP-constrained use cases. Next slide, please.
The PacStar 454, the newest addition to our family, is the higher performance server in our product line equipped with the same Intel processor as the PacStar 453 but this one has an NVIDIA RTX 5000 GPU on it with 3000 CUDA cores and 9.4 teraFLOPS of computing performance. At 210 watts, this is about twice the power and about 33% larger than the 453 but it's a big step up in performance. Next slide, please.
PacStar combines the PacStar 453 and 454 with other modules in our product line along with third-party technologies to create the PacStar Tactical Fusion System, a COTS-based modular tactical and expeditionary rugged processing and data distribution node for sensor-enabled platforms and dismounted warfighters. The system is ideal for processing and analyzing video, invisible and IR spectrums, for target identification, tracking, and handoff, for integrating multiple sensor input from multiple endpoints including from wireless sources. for supporting situational awareness, C2, and mission command applications, distributing data including geographic information and point of interest information to tactical squads, and executing computer vision AI and ML algorithms with low latency. Because PacStar TFS uses the field-proven and highly interoperable PacStar 400 series it offers a vendor-agnostic platform for supporting a vast array of industry-standard devices including COTS and GOTS software and can support any IP-enabled and user devices including cameras augmented or virtual reality headsets and legacy video and radio sources. Last slide, please.
Combined with ancillary equipment, PacStar TFS is the core of a next-generation architecture for processing and data distribution nodes for platform and dismounted warfighters enabling sensor and video-enhanced situational awareness. This full architecture requires a number of classes of components including the node itself comprising general purpose and GPU computing resources, storage, networking and radio gateways, running mission command situational awareness, C2 and management applications, nodal and endpoint radio infrastructure including SATCOM modems, wi-fi devices, and tactical waveforms such as TSM and legacy radios. It includes endpoint devices including video cameras and other sensors, invisible and IR spectrum, and radar and lidar, and endpoint displays including legacy displays, platform displays and warfighter warn AR and VR headsets and mobile devices. While this technology may look futuristic, it's available today based on field-proven COTS technologies from PacStar. So that concludes my remarks. Now I'm turning it back over to Bob Ackerman for our live Q & A.
Bob Ackerman
Okay, thank you, Charlie. Again audience members let me remind you that the ask a question box is your portal to this part of the discussion. Let me begin with some questions we already have. Nima, I think this one is your neighborhood. Why NVIDIA over AMD GPUs?
Nima Neghaban
So, with the NVIDIA GPU and the CUDA language, there was a huge gap in capability. I haven't done a gap analysis where OpenCL is relative to CUDA, but the initial decision was around the CUDA language. With Kinetica, you can leverage two different compute backends:
- one is CUDA backend so if you have an AMD device that supports CUDA 8 or higher you can use an AMD device and
- the other is CPU AVX as a compute backend.
So any processor past 2015 Skylake has AVX 512 capability and can be an option for Kinetica as well.
Bob Ackerman
Okay. Charlie, I think this is your neighborhood. You mentioned your system has a UPS built-in. What is it?
Charlie Kawasaki
Thanks Bob. Nima, if you stop sharing your screen, I can actually show it to you. I have a physical embodiment of it here. Great. Okay so we have several different approaches to UPS and this is an actual PacStar TFS that's sitting behind me. It starts first with a built-in UPS that's in our chassis, that can take wide range AC and DC and has built-in DoD standard 2590 lithium-ion batteries or also options for for nickel metal hydride as well. So, that's built in and it can provide UPS services for the entire set of equipment. The other thing that we have with the PacStar 400 series is, I pulled out one of the modules, you can see it's a modular system and on the side, what we have are radio battery connectors which is a patented solution that allows you to connect standard LAN mobile radio batteries from radio types like the AN/PRC-1148 or 152 and you can snap them right on. You can run these modules that way or you can even snap these modules together and run them off of standard military radio batteries which means that as a program, you don't have to worry about dealing with yet another kind of battery type.
Bob Ackerman
Okay, thank you, Charlie. Nima, I think this question is for you. How is this solution different from other databases that claim to leverage vectorization for enhanced performance?
Nima Neghaban
Yeah, it's a great question, so you may see other databases claiming that you know they're leveraging vectorization but the key differentiation is that you know because we were built from the ground up, all of our database kernels and data structures were built from the ground up to leverage CMD processing. Our complete suite of kernels is optimized where you know pretty much 99% of the field only has small parts of their processing kernels optimized to leverage vectorization.
Bob Ackerman
Okay. Charlie, this one is definitely yours. What other AI applications run on PacStar 453?
Charlie Kawasaki
So, we use industry-standard Intel processors and NVIDIA GPUs, and the GPUs are connected with a very standard enterprise architecture - they're PCIe connected - so application developers who have already developed their applications to take advantage of that architecture in an enterprise context should work just fine on our system. Now, this is a brand new suite of equipment so we are just in the process of qualifying infrastructure software and video analytics software to run on the platform and Kinetica is one of the very first applications that we're announcing supported on the PacStar 400 series. Some other options that we have done preliminarily have been more focused on video transcoding, video analytics so we've done some work with digital barriers and also with Force Point's high-speed guard with video transcoding on board. But Kinetica is the very first database and sensor-oriented application that's running on PacStar 400. I would say we're on the lookout for more so if you have those types of applications that need the right platform to be deployed out at the tactical edge, you know we'd love to collaborate with you.
Bob Ackerman
Okay. Nima, I'm gonna toss this one in your direction. What are we giving up when the data structure is optimized for real-time integration and use? Is the optimization removing other benefits of the original structures such as storage? Are you on mute?
Nima Neghaban
Sorry about that. Actually, with Kinetica, it's the opposite. Usually with sensor feeds what you see happening is roll-ups occur such that you lose the need to have to apply as much compute power, for downstream analysis. With Kinetica, because we're actually able to leverage all of the compute, leverage modern compute we actually keep, we actually allow the analyst or the developer to build very complex aggregates, complex analytics from the granular raw readings themselves. So, we're not doing any kind of you know modification or amelioration of the data where we're doing a roll-up or any kind of intermediate state. We'll bring in the values raw and at their full fidelity and then use compute being able to do those through those fusion operations, do those aggregate operations, and do those on the fly in a performant manner.
Bob Ackerman
Okay, thank you. Charlie, this one I think is hardware-oriented. This looks rugged but what do you mean when you say that?
Charlie Kawasaki
Well thanks, Bob. So that's something that we're super proud of. Each of the individual modules that are in a PacStar 400 series solution go through something like 12 different MIL-STD qualifications for 810 and that's vibration, shock, temperature, blowing sand and dust. We don't just say that we design to those standards but we take them through those procedures with a third-party accredited lab and we can make those reports available. And so, you know any mechanical engineers on the webinar will know there's a big difference between designing to those standards and meeting those tests. We also go above and beyond in that we go through those procedures with the equipment running and that's something that's not necessarily even required. We also do 8 different MIL-STD-461 EMI certifications against every single module. I think we've done over 250, maybe 300 tests now on these solutions, and in some cases we also have special requests from programs that ask us to take systems as a whole through system-specific testing. For example, one of our systems is flight-qualified on the MV-22 Osprey for example and so it's met all the air-worthiness safety certifications as a system. The other thing is, you know as part of Curtiss-Wright, we have other parts of our product line, in some cases with very similar electronics but designed for different use cases and these would be the devices used for things like wash-down requirements where you point a hose at it and spray and in those you'll see the big MIL-999 connectors on some really robust packaging that it looks like you could drive a tank over them, although I don't think that's part of their testing procedure. So now as part of Curtiss-Wright, we have the ability to provide a full suite of different kinds of tactical use cases whether it's dismounted soldier and backpack all the way up through mounted outdoors kinds of applications.
Bob Ackerman
Okay, thank you. Nima, this question I think is oriented toward you. If not, just toss it over to Charlie but is this a product only for the edge, in near edge, or can it be used in other ways?
Nima Neghaban
Yeah, so Kinetica is fully available in the cloud, we have an Azure-managed marketplace application. Our AWS marketplace application is going to be out next month. You can go to our website and start up our developer edition with a simple copy/paste. Kinetica can be used in a variety of form factors and certainly, can be used in the kind of more standard data center or cloud as a managed service. And it can also be used in the near edge in a very small form factor with a small power footprint and develop and deliver a lot of capability.
Bob Ackerman
Let me jump in with my own follow-up to that. Do you have to make major changes or significant alterations to go from the edge to the near edge or vice versa?
Nima Neghaban
No, it's all the same Kinetica experience and that experience has been crafted to meet developer expectations so no exotic data models, no new query languages that you need to learn, it's standard data models using relational idioms and using SQL and a full suite of drivers and connectors to leverage your existing tools.
Bob Ackerman
Okay, all right. This one I'm sure is yours, Charlie. What makes your GPU server different than some of the others on the market?
Charlie Kawasaki
So, several things Bob. First, we've seen a few rugged versions of the NVIDIA Xavier product line that are out in the market today, and while that's really important in places for those products in the marketplace because ours use Intel processors with the PCI connected NVIDIA GPU that means that most enterprise developers can use their existing code on our platform without modification, so it becomes a much more general-purpose solution without asking software developers to recode. The other thing is from a physical perspective our 453 and 454 are much smaller than many of the other products that are out on the marketplace and that's something that we take a great deal of pride in when we're talking about applications that require mobility and if you have to carry something, the smaller that is and the less that weighs, you know our customers significantly appreciate that difference.
Bob Ackerman
This question discusses both hardware and software so I'm going to toss it up and either both of you have at it. Recently there's been a push in DoD requirements for hardware and software that can operate in degraded, intermittent, and low-bandwidth environments. Can this package operate in these environments?
Charlie Kawasaki
I'm happy to take that and then if you want to go Nima. Okay great, so operating in disconnected and degraded environments is exactly what PacStar is all about, that is our business and so everything from being extremely power efficient so that we require less fuel, less batteries to run on, it's something that we have worked tirelessly since the founding of the company to address. We're always looking at processors that are specifically designed to reduce power consumption. When we look at things like GPU-enabled systems, what it provides is a much better sort of compute for your power budget and that's where you know the combination of our hardware with Kinetica technology that takes advantage of that compute resource. You can get a lot more done with your bang for the buck by using some of these new compute architectures. So I think that kind of answers it. Nima, do you have anything else to add?
Nima Neghaban
No, yeah I think that answers it, I mean for Kinetica again you know we're just leveraging that modern computing capability that's being delivered by the GPU so it is the same experience and performance and capability in a disconnected state.
Bob Ackerman
Okay, thank you. Nima, this is your neighborhood again. Okay, do you use any type of ontological framework to arrange data info? If so, is an organization dictating this?
Nima Neghaban
No, that's up to the discretion of the user. We do provide certain data primitives like the track data type that allows you to kind of organize sensor readings so that's basically, an attribute map of sensor readings along with an x, y, and timestamp and a grid that identifies that track. So, we give these primitives to the developer but as far as how you want to organize your schemas and your data sets. we leave that as an exercise to the user. What we do is make it very, dead simple to query an entire schema or join two different data sets across space and time, but how the developer wants to organize that or if they want to dictate like an ontological framework, we have seen customers do that but we don't have any requirements for that to make use.
Bob Ackerman
Okay, thank you. Well, Charlie, we're back in the rugged territory with this question. Are the PacStar chassis designed to be somewhat future-proof? Meaning will modules that come out in 3-5 years from now be able to be swapped in easily and on the fly, hot-swap, and is there redundancy in the chassis?
Charlie Kawasaki
So, great question and let me try to break it down into a couple of parts. So the PacStar chassis accepts standard-sized PacStar modules and as you saw in our slide deck we have a big family of modules today but we are constantly upgrading these modules. If you go and look at our website, for example, you can see we have prior generation routers and new generation routers and we have a whole series of servers that have tracked with advancements in new generations of CPUs and things like that. Because of the way we've designed this system, our customers have been able to simply take out a module and slide a new one in and upgrade that capability so it's already the chassis itself has already proven itself to be highly future-proofed in terms of our ability to upgrade specific point technologies. On the other side of things, PacStar is also highly involved with the DoD and the industry standards organizations that are evolving into looking at future physical form factors and so we're certainly tracking that and making sure that as we continue to develop the platform we can address those kinds of requirements. A really great example is the unit that you see behind me is actually the size of what's called a save enclosure - the standard a kit vehicle enclosure which is an army vehicle standard for where they have a space claim for this kind of equipment. This is a brand new solution for us that is I think a demonstration that we continue to advance the platform to make sure that as our customers change their standards that we can provide this kind of technology to meet those kinds of requirements.
Bob Ackerman
Okay, Nima, one for you. How are you collecting data at the edge or ingesting data from other endpoint devices, network devices IoT devices, that sort and landing the data on your platform?
Nima Neghaban
Right, yeah, that's a great question so Kinetica itself doesn't. It doesn't do kind of raw binary consumption and transformation so customers will leverage other processes that are consuming the raw binary feed to do the transformation to application-level structures like JSON formats, like JSON or Parquet and land those in you know, kind of industry-standard data containers like an S3 bucket, a Kafka topic, an Azure blob store, and we also provide standard drivers like JDBC drivers and ODBC drivers that allow those customers to plug in Kinetica's data consumption-ability into existing ETL pipelines and parsing pipelines that they may have set up to collect directly from those devices.
Bob Ackerman
Okay then. Well, that concludes our Signal media webinar for today. I want to thank our experts for the presentation and thanks to all of you for joining us. Our presenters will try to respond directly by email to any unanswered questions in the queue you can link to the archived version of this webinar along with previous Signal webinars on the Signal magazine website at www.afcea.org/signal/webinar. Thank you again and have a good day.