Chinese researchers use low-cost Nvidia chip for hypersonic weapon —unrestricted Nvidia Jetson TX2i powers guidance system

News
By
published

1.26 FP16 TFLOPS is good enough for fluid dynamics.

Nvidia
(Image credit: Nvidia)

A team of researchers in China has demonstrated how a low-cost Nvidia Jetson TX2i system-on-chip for industrial applications can significantly enhance capabilities of hypersonic weapons — achieving speeds above Mach 7, according to a report by the South China Morning Post. This SoC, which is not restricted by U.S. export controls and is much more affordable than high-end options, could enhance China's military capability. 

The Nvidia Jetson TX2i processor (featuring two Denver 2 cores, two Arm Cortex-A57s, and a Pascal GPU with 256 CUDA cores) was originally designed for industrial applications but has been adapted by the research team to work within air-breathing hypersonic aircraft. These aircraft benefit greatly from the chip's ability to process complex computational fluid dynamics models. This processing power reduces the time required for crucial calculations — from seconds to just 25 milliseconds. 

Economically, using the TX2i module means a significant reduction in research and development costs for hypersonic technologies. With a peak performance of 1.26 FP16 TFLOPS at up to 20W, the TX2i offers substantial computational power at a fraction of the cost of Nvidia's more powerful AI chips. 

The research also highlighted the usage of heterogeneous CPU and GPU computing, presenting a novel architecture for tackling the challenges associated with the sequential nature of hypersonic flow field simulations. This solution is crucial, as it allows for more efficient handling of the complex computations required by such performance hungry applications. 

This advance in computational performance is crucial for 'real-time optimization' of various systems within the aircraft, such as the fuel supply and fault diagnosis in scramjet engines. Such capabilities were detailed in a study that was jointly conducted by the Beijing Power Machinery Research Institute and Dalian University of Technology and published in the Chinese academic journal Propulsion Technology. 

Despite the potential of using a U.S.-manufactured chips like the TX2i, the team acknowledged that Chinese domestic chips could equally meet the demands of the country's military applications — suggesting a possible shift toward greater self-reliance in critical military technologies. This development is a clear indicator of strategic importance of integrating advanced computing solutions, capable of challenging conventional defense systems, in modern weaponry.

Anton Shilov
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

13 Comments Comment from the forums
  • The Historical Fidelity
    Not really surprising, the U.S.’s Sprint anti-ballistic missile designed in the 60’s for the Sentinel program could accelerate from 0 to Mach 10 in 5 seconds (100g’s of acceleration) and was designed to intercept incoming warheads between 15,000-30,000 feet. Sprint was the failsafe quick reaction segment of Sentinel for when warheads survive the main component of sentinel called Spartan, which was designed to intercept and destroy warheads in space via an enhanced radiation nuclear bomb. Needless to say, the failsafe Sprint interceptors had to be exceptionally quick in both acceleration and calculating intercepts in real time as the flight time of an incoming warhead from entering the atmosphere at Mach 25+ to hitting the target was only a few seconds. The point of this story is that the Sentinel system did all of this with 60’s transistor tech, so it’s not really that hard to believe that an Nvidia processor 1000x more powerful than the computers used in the sentinel program can perform similar tasks for the newer/slower Mach 7 missile designed to hit stationary or relatively slow targets.

    P.S. the sprint missile accelerated so fast at such a low altitude (IE thick air) that the missile would glow white hot all the way to interception. Check out the video I linked.
    msXtgTVMcuAView: https://m.youtube.com/watch?v=msXtgTVMcuA
    Reply
  • tanon
    I think you're missing the point here: the Nvidia chip isn't being used to calculate a simple ballistic intercept trajectory (which is actually a fairly trivial calculation that can be done in a few lines with paper and pencil), it's being used to perform fluid dynamics simulations which don't have simple closed form solutions and require complex, iterative calculations to solve (think perhaps a few magazines worth of pages of calculations - which need to be recalculated every few ms as the parameters change)
    Reply
  • The Historical Fidelity
    tanon said:
    I think you're missing the point here: the Nvidia chip isn't being used to calculate a simple ballistic intercept trajectory (which is actually a fairly trivial calculation that can be done in a few lines with paper and pencil), it's being used to perform fluid dynamics simulations which don't have simple closed form solutions and require complex, iterative calculations to solve (think perhaps a few magazines worth of pages of calculations - which need to be recalculated every few ms as the parameters change)
    I think you’re missing the point of my comment. Continuously calculating then recalculating the intercepts of dozens of 0.5 m^2 Mach 25 objects while discriminating in real time warhead decoys with a total intercept window of 2 to 3 seconds from ignition while calculating the fluid dynamic loads and stresses for every minute ABM course correction while traveling at Mach 10 to prevent the missile from disintegrating in flight, while also computing complex algorithms to compensate and filter for surface plasma induced radio interference to maintain telemetry and ground radar feeds is much harder than anything this new missile does. It’s child play in comparison. And this was done 60 years ago….
    Reply
  • FoxtrotMichael-1
    The Historical Fidelity said:
    I think you’re missing the point of my comment. Continuously calculating then recalculating the intercepts of dozens of 0.5 m^2 Mach 25 objects while discriminating in real time warhead decoys with a total intercept window of 2 to 3 seconds from ignition while calculating the fluid dynamic loads and stresses for every minute ABM course correction while traveling at Mach 10 to prevent the missile from disintegrating in flight, while also computing complex algorithms to compensate and filter for surface plasma induced radio interference to maintain telemetry and ground radar feeds is much harder than anything this new missile does. It’s child play in comparison. And this was done 60 years ago….

    As an aviation enthusiast and retro computing enthusiast who had never heard of the Sprint missile system, your comments sent me on a bit of a research rabbit hole. I wanted to learn more about such an amazing missile intercept system and try to answer the question that immediately came up in my mind: could it be that a missile system designed in the 1960s could perform the same or more computationally intensive tasks than a modern Nvidia Jetson SOC?

    The TL;DR is this: the Sprint missile system did nowhere near as complex of calculations as the Nvidia Jetson simulating fluid dynamics. The more complex version involves electro-mechanical systems, analog computing, and the differences between historical flight control systems and modern fly-by-wire systems.

    I found this PDF on the subsystems of the Sprint Missile system. Essentially, it details how there are 3 systems that work together to control an intercept using a Sprint Missile: ground-based radars, ground-based computers, and the Missile-Born Guidance Equipment (MBGE). The MBGE and the Autopilot subsystems are the only systems on the missile itself that do any sort of computation at all.

    The Autopilot is an electro-mechanical system, not digital. If you've ever seen any of the older Air Data Computers that utilize advanced analog and mechanical systems to "compute" air speed and other telemetry from old jets then you'll have the right idea. Think gyros and mechanical springs, not digital circuits. So this leaves us with the MBGE to do any sort of digital computation at all. What did the MBGE do, exactly? From the PDF:
    To receive and decode missile command steering orders
    To receive and decode discrete commands for payload activation, destruct signals, or other purposes
    To receive and decode an autopilot gain control signal which is a function of computed dynamic pressure
    To transpond a beacon signal for ground station radar tracking purposesThe only part of this that was digital, at all, was the decoder: "The decoder employs transistorized digital techniques for message decoding, storage, and signal conditioning."

    So, to be clear, all of the computation for intercept was done on the ground and sent to the Sprint via radio. The engineering behind the Sprint was very impressive! You mentioned one of the more interesting challenges here:
    computing complex algorithms to compensate and filter for surface plasma induced radio interference
    However, even that was done with an analog circuit that amplified and chose the strongest signal. There was no "complex computing algorithm" involved. They did this with good old fashioned electronics ingenuity. The only digital computation performed on the Sprint at all was decoding, storing, and conditioning those intercept signals. We are comparing a digital signal decode filter circuit to an entire computer performing a fluid dynamics simulation in real time. The Nvidia Jetson is multiple orders of magnitude more capable and complex than what was on the Sprint missile.

    So what's the takeaway from all this? For one, modern military hardware with fly-by-wire systems require vastly more computing power for their flight control systems to perform essentially the same job as legacy electro-mechanical analog systems. The upside is that modern systems can push performance envelopes much further by having that real-time simulated fluid dynamics data.

    Anyway, I truly find this subject fascinating and am not trying to prove anyone wrong. Comparing the Sprint Missile to a modern hypersonic missile system is just comparing apples and oranges.
    Reply
  • ivan_vy
    The Historical Fidelity and FoxtrotMichael-1, thank you both, that was an amazing piece of history and engineering insight.
    Reply
  • The Historical Fidelity
    FoxtrotMichael-1 said:
    As an aviation enthusiast and retro computing enthusiast who had never heard of the Sprint missile system, your comments sent me on a bit of a research rabbit hole. I wanted to learn more about such an amazing missile intercept system and try to answer the question that immediately came up in my mind: could it be that a missile system designed in the 1960s could perform the same or more computationally intensive tasks than a modern Nvidia Jetson SOC?

    The TL;DR is this: the Sprint missile system did nowhere near as complex of calculations as the Nvidia Jetson simulating fluid dynamics. The more complex version involves electro-mechanical systems, analog computing, and the differences between historical flight control systems and modern fly-by-wire systems.

    I found this PDF on the subsystems of the Sprint Missile system. Essentially, it details how there are 3 systems that work together to control an intercept using a Sprint Missile: ground-based radars, ground-based computers, and the Missile-Born Guidance Equipment (MBGE). The MBGE and the Autopilot subsystems are the only systems on the missile itself that do any sort of computation at all.

    The Autopilot is an electro-mechanical system, not digital. If you've ever seen any of the older Air Data Computers that utilize advanced analog and mechanical systems to "compute" air speed and other telemetry from old jets then you'll have the right idea. Think gyros and mechanical springs, not digital circuits. So this leaves us with the MBGE to do any sort of digital computation at all. What did the MBGE do, exactly? From the PDF:
    To receive and decode missile command steering orders
    To receive and decode discrete commands for payload activation, destruct signals, or other purposes
    To receive and decode an autopilot gain control signal which is a function of computed dynamic pressure
    To transpond a beacon signal for ground station radar tracking purposesThe only part of this that was digital, at all, was the decoder: "The decoder employs transistorized digital techniques for message decoding, storage, and signal conditioning."

    So, to be clear, all of the computation for intercept was done on the ground and sent to the Sprint via radio. The engineering behind the Sprint was very impressive! You mentioned one of the more interesting challenges here:

    However, even that was done with an analog circuit that amplified and chose the strongest signal. There was no "complex computing algorithm" involved. They did this with good old fashioned electronics ingenuity. The only digital computation performed on the Sprint at all was decoding, storing, and conditioning those intercept signals. We are comparing a digital signal decode filter circuit to an entire computer performing a fluid dynamics simulation in real time. The Nvidia Jetson is multiple orders of magnitude more capable and complex than what was on the Sprint missile.

    So what's the takeaway from all this? For one, modern military hardware with fly-by-wire systems require vastly more computing power for their flight control systems to perform essentially the same job as legacy electro-mechanical analog systems. The upside is that modern systems can push performance envelopes much further by having that real-time simulated fluid dynamics data.

    Anyway, I truly find this subject fascinating and am not trying to prove anyone wrong. Comparing the Sprint Missile to a modern hypersonic missile system is just comparing apples and oranges.
    That’s true, in my mind I was thinking about the basement under the giant radar pyramid with racks of computers when talking about the computations. Re-reading my comment, I did not make that clear.

    https://upload.wikimedia.org/wikipedia/commons/thumb/1/1f/SRMSC_MSR_HAER_ND-9-B.jpg/2560px-SRMSC_MSR_HAER_ND-9-B.jpg
    Reply
  • tanon
    The Historical Fidelity said:
    I think you’re missing the point of my comment. Continuously calculating then recalculating the intercepts of dozens of 0.5 m^2 Mach 25 objects while discriminating in real time warhead decoys with a total intercept window of 2 to 3 seconds from ignition while calculating the fluid dynamic loads and stresses for every minute ABM course correction while traveling at Mach 10 to prevent the missile from disintegrating in flight, while also computing complex algorithms to compensate and filter for surface plasma induced radio interference to maintain telemetry and ground radar feeds is much harder than anything this new missile does. It’s child play in comparison. And this was done 60 years ago….
    The fact someone could seriously believe that a missile from 60 years ago performed real-time fluid dynamics calculations, SDR and pattern recognition functions, at a time when integrated circuits were only 10 years old, is just baffling.
    Reply
  • The Historical Fidelity
    tanon said:
    The fact someone could seriously believe that a missile from 60 years ago performed real-time fluid dynamics calculations, SDR and pattern recognition functions, at a time when integrated circuits were only 10 years old, is just baffling.
    Okay guy, I said the Sentinel system performed fluid dynamics calculations. The sentinel system consisted of the missile, several ground based radars, and a warehouse sized computer mainframe using the cutting edge of computing power from the early 70’s (the missile was designed in the 60’s but the entire system finished design in 74) that performed the calculations and transmitted the resultant guidance information to the missile. Fluid-dynamics calculations used included PMARC (later released for general use as CMARC), USAERO, and the first generation of the still used VSAERO code bases.
    Reply
  • Dicco
    Admin said:
    Nvidia's Jetson TX2i system-on-chip is good enough for hypersonic weapons.

    Chinese researchers install low-cost, unrestricted Nvidia Jetson TX2i into hypersonic weapon : Read more
    Just in case people miss the obvious.

    Anything that is publicly broadcast about what is happening at the cutting edge of military applications is either misinformation, propaganda or a lie.

    Hopefully readers here are smart enough to understand why.
    Reply
  • tanon
    The Historical Fidelity said:
    Okay guy, I said the Sentinel system performed fluid dynamics calculations. The sentinel system consisted of the missile, several ground based radars, and a warehouse sized computer mainframe using the cutting edge of computing power from the early 70’s (the missile was designed in the 60’s but the entire system finished design in 74) that performed the calculations and transmitted the resultant guidance information to the missile. Fluid-dynamics calculations used included PMARC (later released for general use as CMARC), USAERO, and the first generation of the still used VSAERO code bases.
    I have to admit that is a neat trick that I wouldn't have thought of myself: using CFD algorithms that weren't introduced until the 1980s, to perform the fluid-dynamics calculations for a 1970s ABM system, using 1970s computer systems, 10 years prior to those algorithms being available...
    Reply