Pretty sure (but not double checked) that some of this may have been already posted.
Jan '23 NASA SBIR solicitations.
There are a couple of points in here I have noticed and highlighted.
Coincidence?
NASA preferred....
such as 22-nm FDSOI.
Section towards bottom of copy speaks of the inhibiting factors for neuromorphic (Loihi, Akida & Tensor) in low SWAP missions...
require full host processors for integration for their software development kit (SDK) that are power hungry or limit throughput.
I'm curious on that as to the full impact scope as on the BRN website it states the below.
- Power Consumption: Can be implemented in mW—no host CPU required to run the network.
Obviously wider technical considerations above my paygrade that maybe
@Diogenese understands better?
Anyway, we are still being referenced in conjunction with Loihi & Tensor being in the same boat and we are now going 22nm FDSOI as well with next evolution of Akida so still some positives imo.
www.sbir.gov
Deep Neural Net and Neuromorphic Processors for In-Space Autonomy and Cognition
Agency:
National Aeronautics and Space Administration
Branch:
N/A
Program | Phase | Year:
SBIR | Phase I | 2023
Solicitation:
SBIR_23_P1
Topic Number:
H6.22
NOTE: The Solicitations and topics listed on this site are copies from the various SBIR agency solicitations and are not necessarily the latest and most up-to-date. For this reason, you should use the agency link listed below which will take you directly to the appropriate agency server where you can read the official version of this solicitation and download the appropriate forms and rules.
The official link for this solicitation is:
https://sbir.gsfc.nasa.gov/solicitations
Release Date:
January 10, 2023
Open Date:
January 10, 2023
Application Due Date:
March 13, 2023
Close Date:
March 13, 2023 (closing in 39 days)
Description:
Scope Title:
Radiation Tolerant Neuromorphic Learning Hardware
Scope Description:
This hardware scope is for embedded radiation tolerant neuromorphic processors and neural net accelerators that provide hardware support for efficient adaptation and learning in the space environment. The adaptation can be deep learning, reinforcement learning, Hebbian learning, or other machine learning paradigms. To qualify, the hardware must be substantially more power-efficient at learning than central processing units (CPUs) and graphics processing units (GPUs) at comparable technology nodes. Efficiency is primarily measured through trillions of AI operations per watt, where an AI operation is typically a multiply-add. The arithmetic precision expected for digital deep learning is BFLOAT 16 or better, hardware proposals for other learning paradigms or analog hardware should justify their level of precision. The hardware needs to be qualifiable for the space environment, encompassing vibration, temperature extremes, RFI, as well as radiation tolerance for lunar, martian, and deep space missions. Radiation tolerance includes total ionizing dose (TID) immunity at or above 50 krad and no destructive latch up. Note that commercial unhardened devices (COTS) are typically rated below 10 krad. Single-event latch up or unrecoverable faults shall be rare outside of solar flares. The hardware shall be designed to detect and recover from most single event effects encountered in the space environment. Specifically, the number of uncorrected errors in the 90% worst-case GEO environment should be targeted for no more than 1×10-5 uncorrected errors per device-day. In the rare event of an unrecoverable error, the hardware shall support fast reboots. The hardware needs to support the large number of write cycles for synaptic values expected during machine learning. Finally, the hardware needs to support neural net inference in addition to machine learning, preferably within an integrated AI paradigm for in situ adaptation during operations.
The innovation, as compared to terrestrial processors, is to incorporate the mechanisms for fault tolerance in an edge processor capable of machine learning with high power efficiency. Some type of redundancy will likely be needed. The reference for Johann Schumann’s incorporation of triple modular redundancy for Loihi is one example mechanism that masks faults, but at the expense of an overall 3x reduction in power efficiency. In a neuromorphic context with stochasticity, innovations for more efficient fault tolerance techniques might be developed.
Expected TRL or TRL Range at completion of the Project: 2 to 5
Primary Technology Taxonomy:
- Level 1 02 Flight Computing and Avionics
- Level 2 02.1 Avionics Component Technologies
Desired Deliverables of Phase I and Phase II:
- Analysis
- Prototype
- Hardware
Desired Deliverables Description:
Phase I deliverables shall include at the minimum hardware simulation at the Verilog level sufficient for proof of concept of throughput, expected energy efficiency, and redundancy mechanisms for radiation tolerance. Detailed simulations or a tape-out at coarser technology nodes would be a preferable Phase I proof of concept.
Phase II deliverables shall include a prototype processor whose fault tolerance is tested in ground facilities including TID and proton radiation.
The prototype processor and its support circuitry shall be suitable to incorporate on an experimental CubeSat mission, in other words, the printed circuit board (PCB) should fit within 10 × 10 cm. The preference is for a prototype processor fabricated in a technology node suitable for the space environment, such as 22-nm FDSOI, which has become increasingly affordable.
The Phase II delivery should include a maturation plan for a ruggedized production processor fabricated at a competitive technology node with high performance metrics, that could be funded through some combination of outside capital and NASA post Phase II programs.
State of the Art and Critical Gaps:
Neuromorphic and deep neural net computing is a broad field with many technology gaps for space avionics. Through previous and ongoing research and development (R&D), especially under this Small Business Innovation Research (SBIR) subtopic, the SOA in neuromorphic processors for space has advanced to include high throughput, low SWaP, and radiation tolerance—but for neural inference only.
Extended space missions need in situ adaptation and learning for autonomy, otherwise Earth operations are continually remotely updating software in response to unexpected and changing conditions. This adaptation, which characterizes biological systems, requires hardware support for machine learning.
Relevance / Science Traceability:
- 02-10 (Radiation-tolerant Neuromorphic Machine Learning Processors)
- 02-11 (Radiation-tolerant High-performance Memory)
- 03-09a (Autonomous self-sensing)
- 04-23 (Robotic actuators, sensors, and interfaces)
- 10-04 (Integrated system fault/anomaly detection, diagnosis, prognostics)
- 10-05 (On-Board "thinking" autonomy)
Scope Title:
Extreme Radiation Hard Neuromorphic Hardware
Scope Description:
There are two primary differences between this Scope, Extreme Radiation Hard Neuromorphic Hardware, and the Scope titled: Radiation Tolerant Neuromorphic Learning Hardware.
First, the processor is required to have greater radiation hardness. The goal is to develop a processor that is capable of operating through solar flares and the trapped radiation belts of planets such as Earth, Jupiter, and Saturn. This capability means, for example, that a lunar mission does not need to incorporate sheltering in place during a solar flare into its concept of operations. A lunar mission could count on the neuromorphic processor for critical phases, such as entry, descent, and landing (EDL), even during unexpected solar flares. It also enables missions to the outer planets and their scientifically interesting moons. In contrast to the first category, the processor needs to incorporate radiation mitigation measures that meet or exceed TID 200 krad and provide reliable embedded computation during solar flares in deep space. In deep space, the radiation flux during a solar flare can exceed 100 times the background radiation flux, and there are many more highly energetic protons and ion species that penetrate shielding—some up to 100 MeV. Specifically, the number of uncorrected errors should be no more than 1×10-3 per device-minute, for the worst 5-minute period of the October 1989 design case flare in CRÈME 96. See the references on space radiation and electronic effects to calibrate this level of radiation hardness.
Second, the processor could be neural inference-only, relaxing the requirements to support in situ adaptation and learning. To qualify, the hardware must be significantly more power-efficient at inference than radiation hard CPUs, GPUs, and field programmable gate arrays (FPGAs) at comparable technology nodes. Efficiency is primarily measured through trillions of AI operations per watt, where an AI operation is typically a multiply-add. The arithmetic precision expected for digital multiplies is Int8 or better, hardware proposals for analog inference should justify their level of precision. The hardware needs to be qualifiable for the space environment, encompassing vibration, temperature extremes, RFI, as well as radiation hardness for lunar, martian, and deep space missions during solar flares. Radiation tolerance includes TID support at or above 200 krad, and no destructive latch up even under the extreme environment of Jupiter and Saturn. Single-event latch up or unrecoverable faults shall be rare even during solar flares, and the hardware shall support fast reboots.
Expected TRL or TRL Range at completion of the Project: 2 to 5
Primary Technology Taxonomy:
- Level 1 02 Flight Computing and Avionics
- Level 2 02.1 Avionics Component Technologies
Desired Deliverables of Phase I and Phase II:
- Analysis
- Prototype
- Hardware
Desired Deliverables Description:
Phase I deliverables shall include at the minimum hardware simulation at the Verilog level sufficient for proof of concept of throughput, expected energy efficiency, and redundancy mechanisms for radiation hardness to single event effects. Detailed simulations or a tape-out at coarser technology nodes would be a preferable Phase I proof of concept. Simulation of radiation performance would enhance Phase I deliverables.
Phase II deliverables shall include a prototype processor whose fault tolerance is tested in ground facilities including TID, proton, and heavy ion. The prototype processor and its support circuitry shall be suitable to incorporate on an experimental GTO (GeoTransfer orbit) CubeSat mission, in other words, the PCB should fit within 10 × 10 cm. In a GTO mission, the CubeSat experiences daily transitions through the Van Allen belts—roughly comparable to the radiation during a solar flare.
The preference is for a prototype processor fabricated in a technology node suitable for the space environment, such as 22-nm FDSOI, which has become increasingly affordable.
The Phase II delivery should include a maturation plan for a ruggedized production processor fabricated at a competitive technology node for radiation hard processors with high performance metrics, that could be funded through some combination of outside capital and NASA post Phase II programs.
State of the Art and Critical Gaps:
Neuromorphic and deep neural net computing is a broad field with many technology gaps for space avionics. Through previous and ongoing R&D, especially under this SBIR subtopic, the SOA in neuromorphic processors for space has advanced to include radiation tolerance but not radiation hardness.
Radiation hardness enables computing during extreme space environment and events such as solar flares. In order for neuromorphic processors to be used during critical mission phases such as EDL that cannot be postponed, a higher level of environmental robustness is needed. This also opens up these processors for missions such as icy moons of the outer planets.
Radiation hardness could be addressed through techniques similar to radiation hardness for general purpose processors, but also through potentially new neuromorphic techniques. For example, Dual Interlocked Storage Cells (DICE) resist bit flips by requiring simultaneous transition of redundant memory elements, thus masking any radiation noise on one element. However, in a neuromorphic context with stochasticity, a more efficient radiation hardening technique might be to mask noise at the neural equivalent level.
Relevance / Science Traceability:
- 02-03 (Radiation-tolerant High Performance General Purpose Processors)
- 02-10 (Radiation-tolerant Neuromorphic Machine Learning Processors)
- 02-11 (Radiation-tolerant High-performance Memory)
- 03-09a (Autonomous self-sensing)
- 04-23 (Robotic actuators, sensors, and interfaces)
- 04-77 (Low SWaP, “End of arm” proximity range sensors)
- 10-05 (On-Board "thinking" autonomy)
- 10-16 (Fail operational robotic manipulation)
Scope Title:
Neuromorphic Software for Cognition and Learning for Space Missions
Scope Description:
This scope seeks integrated neuromorphic software systems that together achieve a space mission capability. Such capabilities include but are not limited to:
- Cognitive communications for constellations of spacecraft.
- Spacecraft health and maintenance from anomaly detection through diagnosis; prognosis; and fault detection, isolation, and recovery (FDIR).
- Visual odometry, path planning, and navigation for autonomous rovers.
- Science data processing from sensor denoising, through sensor fusion and super resolution, and finally output the generation of science information products such as planetary digital elevation maps.
In this scope, it is expected that a provider will pipeline together a number of neural nets from different sources to achieve a space capability. The first challenge is to achieve the pipelining in a manner that achieves high overall throughput and is energy efficient. The second challenge is to put together a demonstration breadboard integrated hardware/software system that achieves the throughput incorporating neuromorphic or neural net accelerators perhaps in combination with conventional processors such as CPUs, GPUs, and FPGAs. Systems on a chip (SOC), could be another demonstration hardware platform. In either case, the neural cores should do the heavy computational lifting, and the CPUs, GPUs, and FPGAs should play a supportive role. The total power requirements shall be commensurate with the space domain, for example, 10 W maximum for systems expected to operate on CubeSats 24/7 and even less wattage for lunar systems that need to operate on battery power over the 2-week-long lunar night.
The third optional challenge is to evolve the neural net individual applications and pipeline through adaptive learning over the course of a simulated mission.
Radiation tolerance and space environment robustness are not addressed directly through this scope. Rather, a provider is expected to use terrestrial grade processors and only after Phase II target radiation tolerant neuromorphic processors potentially developed under Scopes 1 or 2 or from another source. The goal is to achieve space mission capabilities that require system integration of individual neural nets together with minimal overhead conventional software. The continuous mission-long learning complements the capability of Earth operations to adapt software over the course of a mission.
As background, development of individual neural net software is now state of the practice, and a large number of neural net applications can be downloaded in standard formats such as pseudo-assembly level or programming languages such as TensorflowTM (Google Inc), PyTorchTM (Linux Foundation), NengoTM (Applied Brain Research), LavaTM ( Intel Cooporation), and others. Published neural nets for aerospace applications can be found, ranging from telescope fine-pointing control to adaptive flight control to medical support for astronaut health. In addition, there are many published neural nets for analogous terrestrial capabilities, such as autonomous driving. Transfer learning and other state-of-practice techniques enable adaptation of neural nets from terrestrial domains, such as image-processing for the image net challenge, to space domains such as Mars terrain classification for predicting rover traction.
Expected TRL or TRL Range at completion of the Project: 2 to 4
Primary Technology Taxonomy:
- Level 1 10 Autonomous Systems
- Level 2 10.2 Reasoning and Acting
Desired Deliverables of Phase I and Phase II:
- Analysis
- Prototype
- Hardware
- Software
Desired Deliverables Description:
The deliverables for Phase I should include at minimum the concept definition of a space capability that could be achieved through a dataflow pipeline/graph of neural nets and identification of at least a portion of the pipeline that can be achieved with existing neural nets that are either already suited for the space domain or provide an analogous capability from an Earth application. The pipeline should at a minimum be mocked up and characterized by parameterized throughput requirements for the individual neural nets, a description of the dataflow and control flow integration of the system of neural nets, and an assignment and mapping from the individual software components to the hardware elements, and an energy/power/throughput estimate for the entire pipeline. Enhanced deliverables for Phase I would include a partial demonstration of the pipeline on some terrestrial hardware platform. A report that illustrates a conceptual pipeline of neural nets for autonomous rovers can be found in the reference authored by Eric Barszcz.
The deliverables for Phase II should include at minimum a demonstration hardware system, using terrestrial grade processors and sensors, that performs a significant portion of the overall pipeline needed for the chosen space capability, together with filling in at least some of the neural net applications that needed to be customized, adapted, or developed from scratch. It is expected that the hardware system would include one or more terrestrial grade neuromorphic processors that do the primary processing, with support from CPUs, GPUs, and FPGAs. An alternative would be an SOC that incorporates a substantial number of neural cores. The demonstration shall include empirical measurement and validation of throughput and power. Enhanced deliverables for Phase II would be a simulation of continuous in situ mission-long adaptation and learning that exhibits significant evolution.
State of the Art and Critical Gaps:
Neuromorphic and deep neural net software for point applications has become widespread and is state of the art. Integrated solutions that achieve space-relevant mission capabilities with high throughput and energy efficiency is a critical gap. For example, terrestrial neuromorphic processors such as Intels Cooporation's LoihiTM, Brainchip's AkidaTM, and Google Inc's Tensor Processing Unit (TPUTM) require full host processors for integration for their software development kit (SDK) that are power hungry or limit throughput. This by itself is inhibiting the use of neuromorphic processors for low SWaP space missions.
The system integration principles for integrated combinations of neuromorphic software is a critical gap that requires R&D, as well as the efficient mapping of integrated software to integrated avionics hardware. Challenges include translating the throughput and energy efficiency of neuromorphic processors from the component level to the system level, which means minimizing the utilization and processing done by supportive CPUs and GPUs.
Relevance / Science Traceability:
- 03-09a (Autonomous self-sensing)
- 04-15 (Collision avoidance maneuver design)
- 04-16 (Consolidated advanced sensors for relative navigation and autonomous robotics)
- 04-23 (Robotic actuators, sensors, and interfaces)
- 04-77 (Low SWaP, “End of arm” proximity range sensors)
- 04-89 (Autonomous Rover GNC for mating)
- 10-04 (Integrated system fault/anomaly detection, diagnosis, prognostics)
- 10-05 (On-Board "thinking" autonomy)
- 10-06 (Creation, scheduling and execution of activities by autonomous systems)
- 10-16 (Fail operational robotic manipulation)
- .
Event Link: 
