HR001124S0022_ScAN_FAQ_Updated_08_23_2024.docx

Scalable Analog Neural-networks (ScAN) Program Frequently Asked Questions Document Updated on August 23, 2024

1. Question: Does a subcontractor require a CAGE or NCAGE code? Is the subcontractor also obligated to report finances and undergo audits related to the program funding?

Answer: CAGE codes are required with a few exceptions for subcontractors working on unclassified efforts. The information requirements will change depending on the contracting instrument selected for award by the prime in these cases. Please review DARPA Connect for additional information.

2. Question: Can a commercial company be a sub-contractor?

Answer: An organization/company/individual can be on multiple proposals, including a commercial company.

3. Questions: [Sensor Development]

· Are you looking for an end-to-end processing system that directly interface to an analog sensor?

· Do we have to design an image pixel array to go along with the analog processor?

· Will design of an image sensor such as specialized ROIC within the program scope?

Answer: ScAN will neither develop an analog sensor nor an interface to the sensor, such as a specialized ROIC. Rather, analog signals from the sensor will be simulated. Please see the Broad Agency Announcement (BAA) for more details on program-specific requirements.

4. Questions: [Device Technology]

· What technologies are considered for the program, conventional CMOS or any new device technologies?

· Are we allowed to use a CMOS technology that offers a device type that may not be available in all CMOS technology nodes?

· Is In-pixel computing using non-CMOS, e.g. compound semiconductor and heterogenous packaging with CMOS chip, within the program scope?

Answer: The ScAN program does not prescribe technology choices. However, the proposer must explain how the technology will meet the program objectives. Note that certain technologies and device technology development are out of scope - as described in the BAA.

5. Questions: [Foundry Service]

· Are fabrication and demonstration required? If so, at what scale, what technology? Will DARPA provide any assistance in connecting performers with a foundry?

· Are there any limitations to the process node or foundry selection?

Answer: Fabrication and demonstration are required. The scales of demonstration are defined in the BAA. There is no preferred foundry selection or process node. The ScAN program does not plan to provide specific Government Furnished Capabilities (GFC), such as foundry service assistance.

6. Questions: [Target Applications and Datasets]

· Are there any primary targeted applications, algorithms, or use scenarios?

· Should we expect DARPA to provide relevant data sets to performers as inputs to ScAN or will the performers be providing their own sensor data. If the latter, is there any preference for specific spectral bands or other sensor characteristics?

· What type of demonstration are sought for the software and hardware (benchtop) at each phase? Will data sets be provided?

· Will there be a test use case defined at each phase?

· You have only described image processing in the application space, will IR processing be a use case as well? What about others (LIDAR, multi-spectral cameras, etc.)?

· Could the proposal be written to address a specific application, such as image classification or object detection and tracking?

· Is there a pre-defined list of neural networks for Phase 1 and Phase 2 that need to be supported? What degree of flexibility/programmability needs to be supported in the neural network implementation?

· Do we have to use a particular architecture of the neural network? Is it convolutional net?

· Will there be interest in novel AI algorithms, such as neuro-symbolic or hyperdimensional computing, where the algorithms can demonstrate natural robustness to noise in analog hardware?

· There was a lot of discussion about image processing. Are proposals focusing on RFML (ML applied to wideband microwave/RF signals), i.e., applications to RF sensors instead of image sensors in scope?

· How are test examples going to be delivered to the hardware?

· Can you comment on the type of sensing modalities that are of interest? Most of the examples listed were image or vision based. Are other sensing modalities and applications spaces of interest, such as audio, radar, etc.?

Answer: There is no preferred architecture or algorithm. The proposed approaches are required to meet the program objectives as defined in the BAA. For the purposes of test and evaluation, the BAA specifies benchmarking datasets for each program phase. The Government will provide the benchmarking datasets as Government Furnished Information (GFI) to emulate analog sensors. The proposed hardware demonstration will be required to accept electrical current from those nominal sensors as input. Please see the BAA for more details on program-specific requirements.

7. Questions: [Security Clearance]

· Will performers need to have any security clearance or any particular secure facilities?

· When does DARPA expect to release the ScAN BAA and will there be a classified addendum?

Answer: The ScAN program is unclassified and does not expect any classified effort will be required to meet program objectives. Thus, neither security clearances nor secure facilities are expected to be required. The BAA, as released on 22 May 2024, does not include a classified addendum.

8. Question: Would you be open to proposals for Neural Network architecture research that target mixed analog + digital systems?

Answer: The ScAN program seeks development of all analog and hybrid analog approaches that will achieve the ScAN program objectives (Metrics and Milestones), and that are not specifically excluded per the BAA. Note that hardware demonstrations are required.

9. Question: For each proposal, do you have a rough distribution about software and hardware %? Or it is open for any combination?

Answer: There is no preferred distribution of effort between hardware and software.

10. Questions: [Photonics]

· What role does DARPA expect photonics to play in this BAA?

· Would scalable integrated photonic architectures be considered?

· Is ScAN interested in only analog circuits or are passive optical components that perform neural network activities of interest?

· Are active optical approaches coupled with analog backends of interest?

Answer: There is no preferred technology, but certain technologies and device technology development are out of scope as defined by the BAA. While DARPA does not prescribe technology choices, the Government will provide the benchmarking datasets as GFI to emulate analog sensors. The proposed hardware demonstration will be required to accept electrical current from those nominal sensors as input. Please see the BAA for more details on program-specific requirements.

11. Questions: [Non-Volatile Memory]

· Several technologies are listed as ‘out of scope’, including non-volatile memory-based computing. Neural networks do need memory, volatile or non-volatile, to store weights, etc. Could you elaborate a bit more on this?

· Can you clarify what you mean by “Non-volatile memory-based computing” is out of scope? Do you narrow down to conventional MAC with nonvolatile memory or any design that involves nonvolatile memory?

· Are you only interested in SRAM and DRAM based approaches?

· Are all non-volatile memories excluded or is this statement targeted specifically for passive (resistive) non-volatile memories?

· Does the BAA exclude approaches where weights are stored in the computing devices?

Answer: Teams may propose the use of any memory technology to store weights that will achieve the program objectives (Metrics and Milestones). Non-volatile memory (NVM), when used for multiply-accumulate (MAC) or similar functions, is out of scope as defined in the BAA.

12. Question: Can there be an early tape-out, say to a standard CMOS fabricator, in Phase 1a to help decrease the risk of fabrication in Phase 1b and Phase 2?

Answer: Given limited resources and compressed schedule, critical design reviews are required prior to any tape-out as defined by the BAA. Please see the BAA for more details.

13. Questions: [Input Interface]

· Does the program specifically target an electronic chip that processes electrical input and output? Would a system with optical input and electrical output, or electrical input and optical output, also fall within the program's scope?

· Is development of any optical analog pre-processing subsystems in scope, or do you only want innovations in electronics, not optics?

· Will the demonstrations in Phase 1 and Phase 2 be performed using real-time sensor outputs (for example, by integrating with an image sensor) or using “digital inputs stored in a memory”?

Answer: The benchmarking datasets will be provided as GFI to emulate the analog sensor for the purposes of testing and evaluation. The proposed hardware demonstration will be required to accept electrical current as input from those nominal sensors. Please see the BAA for more details on program-specific requirements.

14. Questions: [Operation Range]

· Does the program have size or reliability targets (such as vibration resistance and operating temperature range) for "actionable intelligence in the field"?

· What is the targeted temperature range for operation? Is it the same as for the conventional CMOS (-55 C to 125 C) Answer: The ScAN program does not specify physical size or reliability targets such as vibration resistance. While the ScAN BAA does not specify an operating temperature range, the ScAN program will initially target -20°C ~ +85°C, and further research the applicable range during the effort. Please see the BAA for more information on metrics and milestones.

15. Question: I believe I caught that solutions that would imply new technology templates for foundry processes are discouraged. As in, there’s no room, especially in Phase I, for solutions that try to forge new processes that don’t already exist at available foundries. With that in mind, would solutions that utilize active circuits be welcome if they conform to *existing* ASIC processes and platforms?

Answer: Correct, however, there is no preferred circuit approach or technology platform.

16. Question: Is the application for ScAN video image classification/inference? There are some NN architectures that are extremely good for such applications (for instance in-pixel NN camera architectures) but not necessarily for general purpose NN architectures.

Answer: Please see the BAA for more information on benchmarking datasets.

17. Question: If a proposed solution is more advantageous for >> 100 M parameters but not as advantageous for < 20M parameters, would it be within scope for DARPA ScAN?

Answer: Please see the BAA for more information on model size requirements.

18. Question: Does [NIST’s] superconducting solution require the typical ultra-low temp operation of the essential circuit components? If so, there goes the edge-computing potential.

Answer: As stated in the ScAN program Special Notice, DARPA-SN-24-57, neither DARPA nor the DoD endorses the information, or the organizations contained in the virtual presentation, nor does DARPA or the DoD exercise any responsibility for improper dissemination of the virtual presentation.

19. Question: How important is it to transition to a program or have a customer demonstration?

Answer: Program transition and demonstration is an integral part of the evaluation criteria under section “Potential Contribution and Relevance to the DARPA Mission.” Please see the BAA for more information.

20. Question: Accuracy suggests a classification task. Is localization of objects in the input not required?

Answer: Please see the BAA for more details on program-specified requirements.

21. Questions: [Budget]

· Did we hear correctly that the technology process (foundry and technology node) that we target is up to the proposers and all fabrication costs should be included in the proposal?

· Should all IP costs (eg, DDR controllers) be included in the proposal budget?

· Should the costs of a board level demonstration be included in the budget?

Answer: All costs that are associated or needed to achieve the BAA milestones, including demonstrations, should be included in the proposed budget. The ScAN program does not prescribe technology choices and does not plan to provide foundry service assistance as Government Furnished Capabilities (GFC).

22. Question: During the presentation, it was mentioned to consider classification problem in Phase I and extend to object detection and tracking problem in Phase II. Are there any suggested datasets to demonstrate the proposed capabilities? In the ScAN program, are there any suggested datasets for performance evaluation? This problem ties to the MNIST dataset point mentioned during the talk as well, where we want a sufficiently challenging dataset that goes beyond toy problems.

Answer: Please see the BAA for more details on benchmarking datasets and program-specific requirements.

23. Question: In addition to PVT variations, ADC offset is another challenge in analog inferencing. Are solutions that address that challenge considered within the program scope?

Answer: The ScAN program does not prescribe a specific circuit approach, but the proposer must describe how their chosen approach will meet the program objectives (Metrics and Milestones). Please see the BAA for more details.

Additions on June 18, 2024

24. Question: Are electro-photonic solutions acceptable for the ScAN program or is the ScAN program exclusively looking for purely electronic solutions?

Answer: The ScAN program does not prescribe technology choices except for those listed as out-of-scope technologies and device technology development efforts. However, the proposer must explain how the chosen technology, whatever it may be, will meet the program objectives.

25. Question: Is there any upper limit on the budget for the ScAN program?

Answer: There is no upper limit or cap on proposed budgets. Note that “Cost Realism” is an evaluation criterion, however.

26. Question: When/where an FAQ will be available?

Answer: The FAQ is published under the “Attachments/Links” section of the ScAN BAA page on SAM.gov. A direct link to the FAQ on SAM.gov is posted under the ScAN announcement on the DARPA Opportunities website. Updates to the FAQ will be annotated as such and be posted periodically to these same locations.

SAM.gov: https://sam.gov/opp/c6687528fcc54113a33256ec445cd0a2/view DARPA Opportunities: https://www.darpa.mil/work-with-us/opportunities?oFilter=MTO

27. Question: [Proposers’ Day Materials]

· Can you please provide the SCAN Industry Day Slides? They are not on SAM.

· Is there are video replay of the proposer day of ScAN?

Answer: The slides from, and recording of, the ScAN program Proposers’ Day are posted under the ScAN announcement on the DARPA Opportunities website.

DARPA Opportunities: https://www.darpa.mil/work-with-us/opportunities?oFilter=MTO

28. Question: The BAA indicates that "diverse expertise from across academia, small business, and the defense industrial base research community may be applicable to achieving all program goals." Does it have to be interpreted into the fact that presenting a proposal that involves a large alpha customer/early adopter part of the Defense industrial base would be a key criterion for DARPA?

Answer: The ScAN program BAA does not require participation/inclusion of customers or early adopters. Rather, the proposed teams’ expertise may be sourced from any combination of academia, small business, and the defense industry-based research community so as to form a diverse and comprehensive team to achieve all program goals.

29. Question: We have a few questions that we would like to have clarification on:

· Phase 1 benchmark target: In Section 1.1.3.1 and Table 1, it is mentioned that performers will develop a demonstration system capable of handling 3,000 inputs, with the analog NN demonstration system needing to be equivalent to a digital model of 1M parameters. Our specific question is: What does the "90% accuracy of the state of the art" requirement in Table 2 refer to? Is "state of the art" referring to a model of at least the size specified in Table 1 and similar architecture? e.g. If different models have varying levels of accuracy and parameter counts (>1M), how should we determine which model to use for comparison (accuracy vs number of parameters)? What if the model's accuracy outperforms other models but with significantly different architectures?

· Phase 2 benchmark target and technical challenge: In Section 1.1.3 and Table 2, the Phase 2 benchmark target includes working with a BDD image dataset for object detection and tracking. The BDD100K dataset provides 10 tasks. Could you please clarify which specific tasks are required with the BDD100K dataset? Specifically, are there detailed requirements or expected outcomes for the object detection and tracking tasks that need to be addressed? We want to ensure we understand the nature of the work to be carried out with this dataset Answer:

· Phase 1: There is no preferred or prescribed SoA reference model. It is up to the performer to propose an appropriate model for implementation that will best demonstrate the architecture robust to variations and power efficient to meet the program metrics and milestones, and, also considers scalability toward Phase 2 target.

· Phase 2: The ScAN program plans to utilize the BDD100k MOT set as a baseline and will initially target Multiple Object Detection and Tracking, or similar tasks, for benchmarking.

30. Question: My team has been in conversation with collaborators to develop a proposal for the ScAN program. We have a system concept and architecture that can meet the program goals with the possible exception of the total system power. Our approach leverages superconducting electronics, so a few hundred watts are required to cool the chip to 4K. We were debating whether your requirement for 1W system power includes cooling power. So we wanted to directly ask: will a system leveraging cryogenic circuits be of interest to you for this program?

Answer: The BAA states that the ScAN program “will develop new analog NNs that could interface directly with the analog outputs of conventional sensors and demonstrate a three-orders of- magnitude power reduction over existing solutions.” While the ScAN program does not prescribe technology choices, the proposed approach will need to be able to interface directly with the analog outputs of conventional sensors, such as nominal Color CMOS Image Sensors specified in the Section 1.1.3.1 of the BAA. In addition, the BAA further specifies that “total system power includes all power consumption from the neural network and peripheral circuits, and pre- and post-processing circuits (analog and digital).” Hence, the proposed approach will need to meet the program-specific requirements and to achieve program metrics and milestones as specified in the BAA.

31. Question: Are we allowed to use a CMOS technology that offers a device type that may not be available in all CMOS technologies? We note that this device type is a standard foundry offering that is part of this CMOS technology node. It is a device type, however, that may not be present in all CMOS nodes. Is targeting such a technology acceptable?

Answer: The ScAN program does not prescribe technology choices except for those listed as out-of-scope technologies and device technology development efforts.

32. Question: We plan to respond to the call and submit an abstract. Could you please introduce how to submit the abstract? Is there a website to submit it online?

Answer: Both abstracts and proposals must be submitted using DARPA’s Broad Agency Announcement Tool (BAAT). This can be accessed via the Proposer Instructions: General Terms and Conditions (darpa.mil) page, in the Unclassified Submission Instructions section.

Additions on August 23, 2024

33. Question: Do approaches that are evolutionary with respect to the state of the art fall within the scope of the program?

Answer: No. If your approach is deemed to be only evolutionary (rather than revolutionary) with respect to the state of the art, then it will be rejected as being outside the scope of the ScAN program.

34. Question: Would it be possible to give examples of mature technology nodes?

Answer: The mature technology nodes for analog circuits are typically 22nm or 28nm.