NuNet: Enabling GPU clusters - Scaling&Expansion

[Proposal setup] Proposal title

Please provide your proposal title

NuNet: Enabling GPU clusters - Scaling&Expansion

[Proposal Summary] Budget Information

Enter the amount of funding you are requesting in ADA

100000

[Proposal Summary] Time

Please specify how many months you expect your project to last

8

[Proposal Summary] Translation Information

Please indicate if your proposal has been auto-translated

No

Original Language

en

[Proposal Summary] Problem Statement

What is the problem you want to solve?

Inability to use multiple GPUs of multiple computers across NuNet’s network hinders many decentralised compute applications, including large scale AI model training, federated learning and others.

[Proposal Summary] Supporting Documentation

Supporting links

• https://nunet.io/
,
• https://gitlab.com/nunet
,
• https://gitlab.com/nunet/device-management-service

[Proposal Summary] Project Dependencies

Does your project have any dependencies on other organizations, technical or otherwise?

Yes

Describe any dependencies or write 'No dependencies'

Internal dependencies (existing NuNet technologies that will be augmented during the project) The primary dependency is NuNet's device-management-service (DMS) protocol, which provides the foundational decentralized orchestration layer for compute resources distributed across globally available compute devices. DMS enables distributed task execution across a network of participating devices, handling resource discovery, allocation, and coordination in a trustless environment. The NuActor system is an integral part of device-management-service which delivers a specialised programming model designed for secure development of decentralised systems. NuActor framework provides isolation guarantees, message-passing semantics, and fault-tolerance mechanisms essential for distributed application reliability. The system abstracts the complexities of distributed state management while maintaining security properties across network boundaries. In conclusion, the majority of the work for this project will go into the enhancement of the Device Management Service to allow job deployments that require a clustered set of GPUs.

[Proposal Summary] Project Open Source

Will your project's outputs be fully open source?

Yes

License and Additional Information

NuNet licensing policy, adopted in April 2023, follows the principle of "Open Source code for community and business use" and publicly available at https://docs.nunet.io/docs/nunet-licensing-policy. It defines a licensing application for all NuNet's projects, publicly hosted at https://gitlab.com/nunet. All components developed as a result of this project will be categorised as core or non-core protocol components and licensed with Apache 2.0 open source software license from the start.

[Theme Selection] Theme

Please choose the most relevant theme and tag related to the outcomes of your proposal.

AI

[Campaign Category] Category Questions

Describe what makes your idea innovative compared to what has been previously funded (whether by you or others).

NuNet’s decentralized infrastructure currently limits GPU utilization to one GPU and one machine, restricting scalable workloads. Our innovation addresses this critical challenge by enabling dynamic, clustered GPU processing across a network of diverse machines. We’re delivering a truly scalable solution, effectively pooling computing resources to maximize performance and bandwidth efficiency regardless of different capabilities. This overcomes the current limitations, offering a robust and adaptable platform for demanding GPU workloads.

Describe what your prototype or MVP will demonstrate, and where it can be accessed.

The prototype will be built on the Device Management Service (DMS) that makes up the NuNet decentralized computing platform. Once completed, it should be possible to utilize multiple GPUs from multiple machines for a single job by pooling together the resources from each machine into an integrated cluster.

Describe realistic measures of success, ideally with on-chain metrics.

Successfully completing this project will enable support for an ensemble specification or job descriptions for the NuNet platform that allows deploying a job with multiple GPU requirement and leverage GPUs across multiple machines. Additionally, example ensemble specifications, documentation, and tutorials will be provided, making this feature easy to use and accessible for both testers and users.

[Your Project and Solution] Solution

Please describe your proposed solution and how it addresses the problem

NuNet has already established a robust GPU identification and computing capability within the Device Management Service (DMS), developed through funding from Cardano Catalyst Fund 8. This initial implementation successfully supports single-GPU jobs, providing reliable resource discovery, scheduling, and execution across the decentralized network. Building on this foundation, the next step is to extend DMS and the existing ensemble specification to enable Orchestrators to run multi-GPU jobs across multiple Compute Providers, unlocking greater scalability and efficiency for high-performance workloads. By developing mechanisms to orchestrate GPU resources across heterogeneous devices in a decentralized environment, we aim to unlock a new layer of scalability and efficiency for NuNet users and contributors alike.

At the core of the solution is the extension of the ensemble specification used for job descriptions to request and coordinate multiple GPUs in a flexible and transparent manner. This extension of the ensemble spec will involve adding new parameters for GPU requirements, such as the number of GPUs, minimum memory per GPU, CUDA/ROCm compatibility, and interconnect performance. For example, a job may specify that it requires four NVIDIA A100-class GPUs or eight consumer-level GPUs with at least 8GB VRAM each. The platform’s scheduling system should then identify available GPUs across the network that collectively satisfy the requirement, regardless of whether they reside on a single machine or multiple compute providers’ devices.

A key technical challenge is enabling distributed GPU execution across machines that may be geographically dispersed and connected over variable network conditions. To address this, the solution will integrate existing distributed computing paradigms such as data parallelism and model parallelism into NuNet’s decentralized architecture. Data parallelism will allow identical model shards to run across multiple GPUs with parameter synchronization at each training step, while model parallelism will divide different parts of a computational graph across GPUs. We will implement middleware that leverages libraries like Horovod, and gRPC-based message passing to ensure efficient communication between GPUs, even in decentralized settings. Additionally, it should also be possible to allow libraries such as NCCL (NVIDIA Collective Communication Library) to be usable as containerized jobs without such a middleware which currently isn't possible due to the way the ip layer on top of the peer to peer network is implemented to forward ports.

The solution also requires the design of a resource discovery and allocation layer capable of securely advertising GPU availability, benchmarking performance, and verifying compatibility. GPUs differ significantly in performance, driver stack, and software dependencies, so a profiling and benchmarking mechanism will be built into the Device Management Serve (DMS). This will allow DMS’s resource allocation to better match job requirements with available GPU resources, optimizing for latency, bandwidth, and cost efficiency. Where possible, GPUs on the same machine or local network will be prioritized to minimize interconnect bottlenecks, while cross-machine GPU coordination will be handled through synchronization protocols that tolerate network variability.

From a decentralization and trust perspective, the solution will build upon NuNet’s existing sandboxing and containerization technologies. Each multi-GPU job will be packaged in a container or virtualized execution environment with explicit resource access controls. Compute Providers will retain full sovereignty over their devices, with configurable limits for their GPUs to be pooled into cluster jobs.

To make this functionality usable and widely adoptable, we will develop documentation, example ensemble specs, and tutorials demonstrating common workflows. For instance, orchestrators will be able to submit a training job for a large transformer model that automatically distributes computation across multiple compute providers’ GPUs, with the system abstracting away most of the complexity. Similarly, tutorials will show how to specify GPU requirements for simulations or inference tasks without needing to manage hardware allocation directly.

The final deliverable will be a production-ready GPU clustering feature integrated into the NuNet platform, complete with APIs for specifying GPU jobs, a scheduling and orchestration layer, and supporting developer tools. This will empower both early testers and users to harness a truly decentralized GPU cluster at scale, paving the way for NuNet to support the next generation of compute-intensive applications in AI, data science, and beyond.

[Your Project and Solution] Impact

Please define the positive impact your project will have on the wider Cardano community

The development of GPU clustering within NuNet’s decentralized computing platform represents a transformative milestone with wide-reaching impact across technological, economic, and societal domains. By enabling single jobs to utilize multiple GPUs across multiple machines contributed by diverse Compute Providers, this feature unlocks a new dimension of scalability, accessibility, and efficiency in distributed and decentralized computing. The integration of this capability within NuNet’s existing ensemble specification and Device Management Service (DMS) ensures that the innovation builds directly upon the platform’s foundations, strengthening its role as a global marketplace for compute.

Empowering Orchestrators with Scalable Compute

The ability to aggregate GPU resources across heterogeneous machines addresses one of the most pressing challenges faced by Orchestrators today: access to scalable, high-performance infrastructure. Training large machine learning models, running simulations, or performing inference at scale requires clusters of GPUs that are typically available only in centralized data centers operated by major cloud providers. By contrast, NuNet’s GPU clustering functionality democratizes access to these capabilities, allowing Orchestrators to seamlessly harness distributed GPUs owned by independent Compute Providers. This reduces reliance on centralized platforms, lowering costs and fostering greater autonomy for developers, researchers, and enterprises.

Unlocking Economic Opportunities for Compute Providers

For Compute Providers, GPU clustering creates new opportunities to monetize underutilized resources. High-performance GPUs are often idle or underused outside peak workloads, particularly in academic institutions, enterprise environments, or even among individual owners of gaming-class hardware. By enabling these GPUs to be pooled into distributed clusters, Compute Providers can participate in high-value compute markets such as AI training and scientific computing. This creates a decentralized revenue stream, incentivizing broader participation and contributing to the growth and resilience of the NuNet ecosystem.

Advancing Decentralized Infrastructure and Innovation

The impact of GPU clustering extends beyond economics into the realm of infrastructure innovation. By demonstrating that complex, high-performance workloads can be executed efficiently on a decentralized and heterogeneous network, NuNet challenges the conventional assumption that such capabilities require centralized cloud-scale infrastructure. This represents a paradigm shift in how compute-intensive tasks can be distributed, coordinated, and validated. In the long term, it sets the stage for more advanced decentralized capabilities, including large-scale AI model training, distributed simulations for scientific research, and federated approaches to compute that prioritize data sovereignty and privacy.

Impact on the Cardano Ecosystem

NuNet’s NTX token is a native asset on Cardano, and the integration of GPU clustering directly strengthens the Cardano ecosystem. NuNet is actively developing payment systems to be available on the platform and by enabling high-performance compute services within a tokenized framework, GPU clustering creates new real-world demand for NTX and drives utility within the Cardano economy. Orchestrators will require NTX to access decentralized GPU resources, while Compute Providers will be compensated in NTX for their contributions, creating a vibrant, self-sustaining marketplace of value exchange. This deepens liquidity and adoption of Cardano-native assets while showcasing Cardano’s scalability for real-world applications.

Furthermore, GPU clustering opens pathways for collaboration with other Cardano-native projects, such as decentralized AI initiatives, research collectives, and DeFi platforms that require compute-intensive analytics. By anchoring these services to Cardano, NuNet strengthens the ecosystem’s positioning as not just a financial infrastructure but also as a foundation for decentralized compute and AI innovation. This alignment demonstrates Cardano’s ability to host advanced, real-economy use cases, reinforcing its competitiveness in the broader blockchain landscape.

In summary, GPU clustering on NuNet is not just a technical enhancement, but a catalyst for broad and lasting impact. It democratizes access to high-performance compute, creates new economic opportunities, strengthens decentralized infrastructure, supports sustainability, and accelerates innovation. By extending the existing capabilities of the DMS and ensemble spec framework, and by anchoring utility to the NTX token on Cardano, this development delivers benefits both to the NuNet network and the broader Cardano ecosystem.

[Your Project and Solution] Capabilities & Feasibility

What is your capability to deliver your project with high levels of trust and accountability? How do you intend to validate if your approach is feasible?

Illustration of Capacity:

Our organization comes with a history of successfully bringing intricate technology projects to fruition. The pillars of our success lie in our deep-rooted technical understanding, stringent project management practices, and an unwavering focus on transparency and responsibility.

Our team consists of experienced software engineers with a strong track record in building distributed computing and decentralized systems. We have already demonstrated our capability by developing the first iteration of GPU support for single-GPU jobs, which is fully integrated into the NuNet platform. The Device Management Service (DMS) that makes up the NuNet platform is currently live and operational, providing a proven foundation for extending support to multi-GPU, multi-machine workloads. This history of successful implementation underscores our readiness to tackle the challenges of the proposed GPU clustering project.

NuNet is committed to Open Source Software development from the inception. Therefore, all our development and progress is available for public scrutiny at all times as well as open collaboration with the community. We actively invite and work with the community in regards to contribution, usage, work and testing of the platform codebase.

Link: https://gitlab.com/nunet

NuNet licencing policy:

https://docs.nunet.io/nunet-licensing-policy/

Openness and Responsibility:

We have established a robust framework to ensure openness and responsibility in the execution of the project and the management of finances:

1. Elaborate Budgeting: We present an exhaustive budget layout at the start of the project that details the fund allocation across various tasks. This leaves no room for ambiguity regarding the utilization of funds.
2. Periodic Reporting: Regular updates regarding the project and financial statements will be shared, offering complete transparency into the progression of the project and the use of funds.
3. External Auditing: We are open to audits conducted by independent third parties at regular intervals. This ensures responsibility and openness in our financial management.
4. Payment Based on Milestones: Our payment structure is built around specific, agreed-upon milestones. This ensures that funds are released as we achieve these milestones. The completion of each milestone can be verified, ensuring you pay only for verifiable progress.

These measures reflect our commitment to openness, responsibility, and proper management of funds. We believe that these factors, along with our technical capabilities, make us an ideal choice to successfully execute this project.

We understand that not all steps we have implemented are valid for the Catalyst proposal but it demonstrates the internal working procedures we have in place.

Catalyst Experience

NuNet also has received the funding for proposals in Fund7 and Fund8. Both proposals have been delivered. Overall, the funds were spent as intended on the development which can be monitored on Gitlab with daily commits since the award.

https://gitlab.com/groups/nunet/-/milestones/19#tab-issues

https://gitlab.com/groups/nunet/-/milestones/20#tab-issues

Financial Stability

As a 20+ strong team, we have independent funding to develop the core platform with a cash runway for at least 1-1.5 years. Cardano Catalyst proposals are used to extend the functionality and add features to the platform in order to enrich the possible use cases.

Risk analysis and mitigation:

Insufficient Compute Providers – If the network does not have enough Compute Providers available to support multi-GPU jobs due to the substantial resource requirements of high-performance workloads, a compensation structure will be implemented in which Orchestrators pay NTX to Compute Providers for access to their GPUs, similar to a cloud computing model. Additionally, NuNet will deploy a portion of its own GPU resources as a fallback to ensure network balance and maintain availability for critical jobs at all times.

[Milestones] Project Milestones

Milestone Title

Requirements and Architecture Design

Milestone Outputs

NuNet will provide the following:

• Detailed technical specification for multi-GPU, multi-machine support
• Updated ensemble spec extensions in DMS
• System architecture diagrams and workflow documentation

Acceptance Criteria

NuNet will provide the following:

• Architecture addresses distributed GPU scheduling, orchestration, and execution
• Ensemble spec extensions clearly defined
• Stakeholder review and approval completed

Evidence of Completion

NuNet will provide the following:

• Signed-off technical specification document
• Architecture diagrams with annotated workflows
• Review meeting notes or approvals

Delivery Month

1

Cost

15000

Progress

10 %

Milestone Title

Resource Discovery and Allocation Layer

Milestone Outputs

NuNet will provide the following:

• Implementation of GPU resource profiling and identification
• Allocation module capable of allocating multiple GPUs across Compute Providers
• Network-aware allocation logic considering interconnect performance

Acceptance Criteria

NuNet will provide the following:

• Onboarding correctly identifies available GPUs meeting job requirements
• GPU Identification and Benchmarking accurately reflect GPU performance and interconnect conditions
• Pass a simulated multi-GPU test scenarios

Evidence of Completion

NuNet will provide the following:

• Test logs showing successful multi-GPU allocation
• Benchmark reports for multiple GPU types
• Scheduler source code reviewed and approved

Delivery Month

4

Cost

30000

Progress

40 %

Milestone Title

Distributed Execution Middleware

Milestone Outputs

NuNet will provide the following:

• Middleware integration for distributed execution using data and model parallelism
• Communication protocols implemented (NCCL, Horovod, gRPC)
• Containerized job execution environment with GPU access

Acceptance Criteria

NuNet will provide the following:

• Multi-GPU jobs execute correctly across multiple Compute Providers
• Synchronize parameters efficiently and tolerates network variability
• Security and sandboxing enforced in all containers

Evidence of Completion

NuNet will provide the following:

• Successful end-to-end test runs of multi-GPU jobs
• Logs showing correct parameter synchronization and task completion
• Security and containerization audit reports

Delivery Month

6

Cost

30000

Progress

70 %

Milestone Title

Documentation, Tutorials, and Examples

Milestone Outputs

NuNet will provide the following:

• User documentation for Orchestrators specifying multi-GPU jobs
• Example ensemble specifications for common workloads
• Step-by-step tutorials for testers and early users

Acceptance Criteria

NuNet will provide the following:

• Documentation is clear, comprehensive, and reviewed by stakeholders
• Tutorials are executable and demonstrate all major functionalities
• Example ensemble specs cover diverse GPU configurations

Evidence of Completion

NuNet will provide the following:

• Published documentation and tutorial files
• Peer review feedback confirming clarity and usability
• Sample job runs following tutorials

Delivery Month

7

Cost

10000

Progress

90 %

Milestone Title

Milestone 5: Testing, Validation, and Production Release

Milestone Outputs

NuNet will provide the following:

• Full system testing and validation of multi-GPU jobs
• Performance and reliability reports across diverse network conditions
• Production release of GPU clustering feature

Acceptance Criteria

NuNet will provide the following:

• All critical bugs will be resolved and multi-GPU jobs pass functional and performance tests
• Feature integrated with DMS and ensemble spec, ready for Orchestrators

Evidence of Completion

NuNet will provide the following:

• Test logs and validation reports
• Release notes and production deployment records

Delivery Month

8

Cost

15000

Progress

100 %

[Final Pitch] Budget & Costs

Please provide a cost breakdown of the proposed work and resources

The following budget breakdown provides an overview of the anticipated costs essential to delivering NuNet’s GPU clustering functionality. This allocation supports every critical stage, including project management, coordination, software development, testing, and associated research of ecosystem technologies, ensuring that NuNet has the resources necessary to build a robust, production-ready multi-GPU, multi-machine system.

Outlined in ADA, the budget is categorized to cover project management, IT development, system integration, testing, documentation, and research. This structured approach ensures operational efficiency, seamless technical execution, and a strategic reserve for unforeseen requirements — setting a strong foundation for project success.

Milestone #1: Requirements and Architecture Design

15,000 ADA / Resources & Skills Needed: Systems architects, software engineers/developers

Milestone #2: Resource Discovery and Allocation Layer

30,000 ADA / Resources & Skills Needed: Software engineers, DevOps, infrastructure for testing

Milestone #3: Distributed Execution Middleware

30,000 ADA / Resources & Skills Needed: Software developers, distributed computing engineers, testing infrastructure

Milestone #4: Documentation, Tutorials, and Examples

10,000 ADA / Resources & Skills Needed: Technical write ups and documentation, tutorial creation

Milestone #5: Testing, Validation, and Production Release

15,000 ADA / Resources & Skills Needed: QA engineers, performance benchmarking and testing infrastructure

[Final Pitch] Value for Money

How does the cost of the project represent value for the Cardano ecosystem?

This project delivers exceptional value by addressing a critical need for high-performance, decentralized GPU compute. Building on NuNet’s existing single-GPU functionality, it enables multi-GPU, multi-machine workloads, significantly expanding the platform’s capabilities for Orchestrators and Compute Providers alike.

By allowing jobs to leverage GPUs across multiple machines, the project enhances scalability, efficiency, and resource utilization. Orchestrators gain access to flexible, cost-effective compute resources without relying on centralized providers, while Compute Providers are able to monetize underutilized GPUs, creating a sustainable and self-reinforcing ecosystem.

From an operational perspective, the platform reduces risks associated with hardware bottlenecks, single points of failure, and underutilized resources. The decentralized architecture ensures that workloads continue to execute reliably even when individual nodes experience downtime or performance variability.

By combining distributed GPU orchestration, efficient resource utilization, and a tokenized settlement mechanism, this project delivers a robust, production-ready solution that strengthens NuNet’s platform, unlocks new opportunities for complex workloads, and ensures the network remains flexible, resilient, and sustainable.

[Required Acknowledgements] Consent & Confirmation

Terms and Conditions:

Yes