Cloud computing, supercomputers, black boxes, and the Kill Web
BlogNovember 30, 2020
Ray Alderman
VITA Standards Organization
WARFARE EVOLUTION BLOG. Back in 1991, U.S. and coalition forces decimated the Iraqi Army in 42 days during Operation Desert Storm. At the time, Iraq had the world’s fifth largest army. Can we do better than 42 days in the future? Yes, with the help of cloud computing and a supercomputer.
In August of 2019, the U.S. Army bought a 6-Petaflop supercomputer from IBM. About 1,000 high-performance processors, and their memory, power supplies, and storage devices, are enclosed inside a metal cargo container, like the ones used on commercial container ships. The 6-petaflop performance level says it will be doing a lot of calculations, and the cargo container says it can be shipped to any place in the world in a matter of hours by a military transport plane. To give you an idea of how powerful this computer is, a petaflop is one quadrillion floating-point mathematical calculations per second. This machine can accomplish 6,000,000,000,000,000 of those calculations every second.
In October 2019, the Pentagon awarded the JEDI (Joint Enterprise Defense Infrastructure) cloud computing contract to Microsoft. Google dropped out of the bidding early, Amazon protested the contract award to Microsoft (twice), and the deal is still hung-up in legal wrangling. So, how do supercomputers and cloud computing fit into the Kill Web? That’s the question we'll explore here.
Before we step into this technical and operational morass, we need some perspective to eliminate confusion. The primary proficiency of a supercomputer is computation. Thousands of tightly-coupled microprocessors manipulate billions of bytes of data into thousands of bytes of decisions in seconds. The primary proficiencies of cloud computing are storage and retrieval of data. Thousands of loosely-coupled servers manipulate thousands of bytes of inquiries into millions of bytes of responses in minutes. Cloud computing data centers, safely away from the battlefield, do the strategic planning. Supercomputers, near the battlefield, do the tactical execution.
To read more Warfare Evolution Blogs by Ray Alderman, click here.
Let’s take a look at how they work together. In September 2020, the U.S. Army gathered imaging data from LEO [low Earth orbit]satellites, Grey Eagle drones, and combat ground vehicles in the desert at the Yuma Proving Grounds in Arizona during "Project Convergence." All the data from the different platforms and sensors was sent to a cloud computing center at Joint Base Lewis-McChord in Washington state 1,300 miles away, on a radio-frequency mesh network. The images were merged and analyzed by the computers at the cloud data center. An enemy tank, and its ground coordinates, were identified by an artificial intelligence (AI) algorithm named “Prometheus" running on those servers.
Prometheus-AI then sent the target coordinates and composite images of the enemy tank to computers at Yuma Air Station back in Arizona. Those computers were running another AI algorithm: FIRESTORM (FIRES synchronization to optimize responses in multi-domain operations). FIRESTORM-AI integrated weather data, terrain data, and locations of friendly forces into the composite picture. Then, FIRESTORM-AI chose the appropriate weapon, already connected in the mesh network, to fire at the target. That weapon was a 155mm ERCA (extended range cannon artillery) located 30 miles away from the enemy tank. FIRESTORM-AI (at Yuma Air Station) took control of the cannon (somewhere in the desert), autonomously aimed the barrel to the required azimuth (direction) and elevation (arc), and authorized the artillery officer to pull the trigger. From collecting the images generated by all the platforms, through target-ID and coordinate determination by Prometheus-AI, to FIRESTORM-AI choosing ERCA, autonomously aiming the cannon, and authorizing the artillery officer to fire the weapon, took a total of 20 seconds. The artillery round took about a minute to cover the 30 miles (a high arc) and hit the target.
Twenty seconds is impressive, until you understand that the ADV satellite ground terminals use 8-digit grid coordinates while the AFATDS fire control system on the cannon uses 10-digit grid coordinates. A soldier had to manually convert the 8-digit satellite-generated coordinates to 10-digits before they were shipped-off to the Prometheus algorithm. That took at least five seconds. A software patch, to do the conversion automatically, is in the works. We’ll expand on this data-structure problem later.
Generally speaking, the computers running Prometheus-AI at Joint Base Lewis-McChord were the cloud, and the computers at Yuma Air Station running FIRESTORM-AI were the supercomputer. You probably noticed here: this was a very simple wartime scenario with one target and one responding weapon. Now consider a complex scenario: hundreds of enemy targets and hundreds of responding weapons (bombers, fighter planes, artillery, tanks, armed UAVs, missiles on Navy ships, missile-armed helicopters, etc) all connected to the mesh network for FIRESTORM-AI to control. That will require the processing power of a supercomputer, near the war zone, like the cargo-container supercomputer.
Let’s replay the Gulf War from 1991, when Iraq invaded Kuwait, but with the cloud computers and the containerized supercomputer involved. After Saddam's invasion, military planners would start sifting through all the data about Iraq’s army strength, command and control structure, terrain, location of enemy troops and weapons, and weather from the data bases stored in the cloud. They come-up with a master plan to liberate Kuwait. The allied commanding generals start moving combat infantry divisions, artillery, tanks, fighter planes, bombers, and supplies into position in friendly Middle East countries, near the borders of Kuwait and Iraq.
They also move the containerized supercomputer to Aviano Air Base in Italy, task the Air Force-controlled LEO imaging satellites to fly over Kuwait and Iraq, and they start connecting the ISR (intelligence, surveillance, and reconnaissance) drones and aircraft, along with all the deployed weapons, into the mesh network. IMINT, [image intelligence] SIGINT [signals intelligence], and ELINT [electronic intelligence] sensor data is fed from the war zone into the cloud computers running Prometheus-AI back in the states, via satellite communication links. Hundreds of targets are found, identified, fixed, and tracked continuously. This is the strategic planning function of cloud computing.
Prometheus-AI then sends hundreds of target coordinates and composite views of the battlefield to the containerized supercomputer running FIRESTORM-AI in Italy, via satellite links. FIRESTORM-AI begins matching the identified targets to the weapons moving out of Saudi Arabia and the Persian Gulf into Kuwait and Iraq. Then, FIRESTORM-AI takes control of those weapons through satellite links, assigns specific weapons to specific target coordinates, and issues the authorization for the soldiers to fire at specific times (yes, there is an optimum pattern of enemy target destruction, and FIRESTORM-AI has calculated that pattern). This is the tactical execution function of the containerized supercomputer. How long would this battle last? Maybe 48 hours. The Iraqi Army could not retreat faster than that.
Now, let’s use the 5F tactical Kill Web model to look at the timing of the sequence in Project Convergence. The Find (identify), Fix (track), and Fire phases took 20 seconds. The Finish phase (the time for the artillery round to cover the distance and hit the target) took about one minute. The Feedback phase (a circling Reaper drone sending after-attack video of the burning enemy tank back to FIRESTORM-AI) would take another 10 seconds. Total elapsed time: 90 seconds. Can we do better? Yes. Just consider that some of the hardware, software, and communications links involved in Project Convergence were prototypes. Clean-up the AI algorithms and data-conversion software, and they can probably get this entire sequence down to a little over one minute. AF General John Jumper would be proud: his goal was to hit any identified enemy target in 10 minutes or less.
Let’s look at the programs underway, to tie all the services’ ISR and weapons platforms together on the mesh network, and create the Kill Web. A diagram of these elements would look like a pyramid. At the top is the Pentagon’s JADC2 (Joint All-domain Command and Control) program. At the next level, we would see the Army’s IBCS (Integrated Battle Command System), the Air Force’s ABMS (Advanced Battle Management System), the Navy’s CEC (Cooperative Engagement Capability) program, and the JAIC (Joint Artificial Intelligence Center).
Under the Army’s IBCS, we find AFATDS (Advanced Field Artillery Tactical Data System), another network for their ISR drones, another network connecting infantry soldier’s tactical cellphones together (TAK, or Tactical Assault Kit), and some other stuff. Under the Air Force ABMS, we would see tactical drone networks, fighter plane networks, tanker plane and cargo plane networks, air traffic control networks, and some other stuff. Under CEC, we would see the NIFC-CA (Naval Integrated Fire Control-Counter Air) network, fighter plane and tanker plane networks, ship-to-ship networks, submarine networks, and Marine Corps soldier and weapons networks. Under JAIC, we would see Project Maven, Prometheus, FIRESTORM, and hundreds of other AI projects. Many of these existing TDLNs (tactical data link networks) connect to our NATO allies’ platforms with LINK hardware and software (Link-1, Link-4, Link-11, Link-14, Link-16, Link-22), depending on how old or new they are. So, add those into the mix.
As you can see, there are many small TDLNs at the bottom of the pyramid, running on old legacy platforms. They are proprietary, designed as “stovepipes” by each of the services and the different countries years ago, and were never meant to talk to each other. They have different data structures, operate at different frequencies, and use different control sequences. Replacing all that communications hardware, and writing all that new software, would be expensive and time consuming. So, those older systems will be connected to the next higher level in the pyramid with a "black box," that translates the legacy data and control structures to the new data and control structures used in ICBS, ABMS, and CEC. That requires the Army, Air Force, and Navy coming together and defining common data standards. David Spirk is the Pentagon’s new Chief Data Officer, and he has an inter-service committee working on that.
Back in October, the Army and Air Force signed an agreement to combine parts of ICBS and ABMS. That changed JADC2 to CJADC2 (Combined Joint All-domain Command and Control). Project Convergence is now an incorporated effort between the Army and the Air Force. So far, the Army has already conducted three tests under Project Convergence and the Air Force has conducted two tests under ABMS. Obviously, they have already solved some of the data structure and control-sequence problems.
The Navy has generally been a spectator in the Project Convergence and ABMS testing. Under CEC, they are concentrating on connecting all their existing ships, ISR systems, and weapons TDLNs together first, with "Project Overmatch." Once those legacy networks are interconnected, they will plug those into CEC and then CJADC2 with black boxes. The concept of using black-boxes, to interface older systems to new ones, has been around for a long time. Technically, things will be messy in all the services for a while. But, as next-generation platforms are designed, built, and deployed using open systems hardware and software standards (MOSA, SOSA, CMOSS, FACE, VICTORY, etc), the black boxes will go away. Now, you know where all these lower-level technical standards fit in the big picture.
Could I draw a diagram that shows how all the ISR, weapons, cloud computers, and the supercomputer are connected in the mesh network? Yes, but it would only be accurate for a few minutes. Hundreds of ISR and weapons platforms could be entering and leaving the mesh network regularly, to refuel and re-arm. So, the structure of the mesh network is dynamic and constantly changing. Human brains can’t keep up with all these variations, but Prometheus-AI and FIRESTORM-AI can.
At this point, your brain is probably drowning in a sea of acronyms, so let’s get back to dry land. Using a combination of known facts, inductive and deductive logic, and some speculation, I have connected a lot of dots for you here. That’s how I see the Kill Web developing, both technically and operationally. As Project Convergence, ABMS, Project Overmatch, and JAIC testing continues, we’ll have more details and a clearer picture. You really should read about all the programs mentioned here for a deeper understanding.
There is another area that raises some command-and-control (C2) questions about the Kill Web, and that’s Kill-TV." With different levels of commanders in different locations viewing the battle space through live video feeds, who ultimately has the authority to order the firing of the weapons? Some general or a computer? That’s what we’ll chew on next time.
