Gintonic Tech Description

Introduction

Gintonic is an advanced AI deployment platform that simplifies the launch and operation of APIs for open-source AI models. It allows developers to deploy and execute models on a decentralized network of GPUs, with each deployment packaged in a Docker container. These containers provide a fully documented API (via Swagger), enabling seamless integration with existing systems. By leveraging Hugging Face’s AI models, Gintonic ensures that users have access to trusted, high-performance solutions.

Problem

The current landscape of AI infrastructure is dominated by centralized cloud providers like AWS and Google Cloud, which often come with high costs, vendor lock-in, and potential vulnerabilities to system outages. On the other hand, decentralized platforms tend to struggle with performance, scalability, and availability, making them less reliable for handling AI’s computationally demanding tasks.

Why existing projects fall short:

1. Centralized solutions:

• High resource costs with little flexibility in usage conditions.

• Limited customization of AI models for specific tasks.

• Vulnerable to technical failures, which can lead to complete service outages.

1. Other decentralized projects:

• Low performance - many decentralized projects cannot offer sufficient power to handle complex computational tasks.

• Security issues - without proper management, decentralized solutions can face security risks and unreliable providers.

A detailed examination of various platforms for renting graphics processing units (GPUs) has revealed several issues:

• IO.net encountered difficulties when deploying clusters, including delays in launching and accessibility issues through various interfaces. Users frequently had to reload connections, resulting in significant access delays.

• Akash Network experienced a lack of available nodes for task distribution, leading to scheduling failures and lost connections to the control console, limiting troubleshooting capabilities and system status monitoring.

• Netmind identified errors in the GPU rental payment process, where transactions failed to complete and funds were not deducted from user accounts, disrupting the purchasing process.

• Octaspace displayed a complete absence of active nodes on the platform, making it impossible to rent computational power. This limits the platform’s functionality and prevents users from utilizing the offered services.

• Node AI, after successfully processing payments for GPU rental, failed to provide access to the resources. Users faced errors and a lack of information on the operation’s status, questioning the platform’s reliability.

• Runpod.io has issues integrating popular decentralized wallets and site malfunctions when selecting cryptocurrency payments, degrading the user experience and limiting the service’s availability for those who prefer or rely on using cryptocurrencies.

• Spheron exhibits slow connectivity and lacks support for installation on the Windows operating system, creating barriers for Windows users and increasing the complexity of setup for new users.

Why this is important:

As AI advances, the demand for scalable, secure, and accessible computational power grows. Centralized providers introduce cost and trust issues, while decentralized networks have yet to deliver the necessary reliability and performance. Solving this problem is critical to enabling a more open, efficient AI ecosystem that can serve a wide range of users.

Decentrized GPU projects testing:

1. IO.net

GPU mega cluster prices:

Test results from io.net

https://docs.google.com/document/d/1nk25p-VU0P5qE2vtlKnSC1GFXmk4qElN9jlCOdKNCxc/edit

Upon deployment of over 6 clusters I have ran into same issue over and over again. Using Decentalized Ray Cluster of General cluster type with io.net as the cluster supplier, E2E encryption, and testing connection to Singapore, Vietnam, Thailand and US. With Ultra High, High, and Medium speed running GeForce RTX 3070, 3080, 4070, 4080 and 4090 on Ray App and IO Master Starter. I have deployed 1 hour long clusters.

The clusters took around 3 minutes to sprint up to be ready for usage. I got 3 options to access cluster, through VS, Jupiter and Ray, however all 3 links could not load properly. All 3 links were reloaded multiple times, and waited for up to 10 minutes, however none of them were loaded. Only once was I able to deploy a cluster which was accessible through Jupiter onl

2. Akash network

Website: https://akash.network/

GPU renting price

A testing session was conducted for the Akash Network, during which an attempt was made to deploy nodes within the network infrastructure. The primary objective of this test was to evaluate the functionality and stability of the node deployment process, along with ensuring proper resource allocation and connectivity.

• Node Deployment Failure:

  The node failed to start as expected. According to the logs, there were issues related to FailedScheduling due to a lack of available nodes. The system reported that no nodes were available for scheduling the pod, indicating a resource constraint or configuration issue with the deployment environment.

• Persistent Volume Claim Issue:

  There was an issue with the Persistent Volume Claim (PVC) for the pod, which was not immediately bound, leading to a delay in resource allocation. This may have contributed to the node deployment failure.

• Shell Access Problem:

  The connection to the Akash Console Shell was lost, making it impossible to interact with the system or troubleshoot through the shell interface. This lack of shell access further hindered real-time debugging and manual inspection of the deployment environment.

• Pod Container Back-off Error:

  After pulling and creating the container for the service vl1m, the system encountered a BackOff error when restarting the failed container. This indicates that the container failed to run successfully after multiple attempts, potentially due to the earlier PVC issue or misconfiguration.

3. Netmind

Web-site: https://power.netmind.ai/rentIntro?gad_source=1&gclid=EAIaIQobChMIte7ylfO_iAMVcBWDAx0n9xX4EAAYASAAEgIBTfD_BwE

GPU renting price

During an attempt to purchase an NVIDIA GeForce RTX 3090 GPU on the Netmind platform, an issue was encountered where no funds were deducted from either the Bybit or MetaMask wallets. The transaction involved ETH, NTM, and USDC, but none of these funds were processed. Attached is a screenshot showing the issue. The steps to reproduce the problem are as follows: select the NVIDIA GeForce RTX 3090 GPU for purchase, choose either Bybit or MetaMask wallet for the transaction, and attempt to complete the purchase. It was observed that no funds were deducted from the selected wallet. The expected behavior is that funds should be successfully deducted from the chosen wallet for the GPU purchase.

4. Octaspace

Web-site: https://octa.space/

GPU rental price

The Octaspace platform has a dashboard designed to track active online nodes. The project doesn't have its own computer resources, which would be always available for rent. However, as of now, the dashboard displays zero active nodes, meaning that it is currently impossible to rent computational power through the platform.

This absence of available nodes limits the platform’s functionality, preventing users from utilizing the intended services for renting or accessing computational resources.

5. Node ai

GPU price rental

On the NodeAI platform, a GPU RTX 6000Ada was purchased with the minimum payment for a 12-hour usage period. The payment was successfully processed, and the funds were deducted from the wallet. However, an issue was observed one hour after the purchase: the GPU remains unavailable and errors persist, as shown in the attached screenshots.

Despite the payment being completed and funds deducted, there has been no progress in making the GPU operational. The website provides no further information or updates on the status of the GPU.

6. Runpod.io

Web-site: https://www.runpod.io/

Current prices for GPU:

Report on Runpod.io Payment System:

During the testing of the Runpod.io platform, it was found that when attempting to pay using cryptocurrency, the website ceases to function properly. In addition to the absence of integration with popular decentralized wallets like MetaMask, Trust Wallet, or Phantom, the platform becomes unresponsive when cryptocurrency payment methods are selected. This issue prevents users from completing transactions and significantly impacts the user experience, especially for those who prefer or depend on cryptocurrency for payments. Addressing this malfunction would be crucial for improving the platform's usability and accommodating a broader user base.

7. Spheron

Web-site: https://www.spheron.network/

During the process of setting up the Spheron Protocol, several limitations were identified, which may impact the user experience, particularly for those using the Windows operating system. These findings include:

1. Protocol Installation Unavailable on Windows: Currently, Spheron Protocol’s CLI (sphnctl) is only available for installation on Linux and macOS. Windows users are unable to directly install the protocol, creating a significant accessibility barrier for a large portion of potential users.

2. Dependency on Curl Installation: For users on supported operating systems, installing the Spheron CLI requires the presence of curl on the system. This adds an extra layer of complexity for users who may not already have curl installed, requiring them to install this tool before proceeding with the protocol setup.

3. Account Creation Requirement: Prior to utilizing the protocol, users must create an account, which could be an additional inconvenience for some. This registration process, while necessary for wallet creation and protocol interaction, adds further time and steps to the overall deployment process.

4. Slow Protocol Connectivity: Connecting to the Spheron Protocol, especially during the deployment phase, takes an extended amount of time. This can slow down operations, leading to inefficiencies and frustration for users seeking a more streamlined experience.

There is currently one active server that is inaccessible, and no actions can be performed with it, including deletion or modification. Attempts were made to complete transactions using Spheron’s native currency, including using their built-in NMT/USDC conversion, but these efforts were unsuccessful. Following their recommendations, MetaMask was downloaded and installed. The payment was processed successfully, and the funds were deducted, but connectivity to the server was not achieved. Further assistance from Spheron support is required to resolve the server access issue and verify that all transaction-related configurations are correctly set up.

9. GPU.net

Web-site: https://www.gpu.net/

The GPU.net platform is undergoing continuous updates and improvements due to the frequent occurrence of errors reported by its users. According to user feedback, these errors significantly impact the platform's stability and functionality. As a result, users often experience difficulties accessing the platform, with disruptions and downtime being a common issue.

In addition, the constantly evolving protocol of the platform exacerbates the problem, preventing users from utilizing GPU.net effectively. The frequent updates, while intended to address these issues, often lead to further complications, making it difficult for users to rely on the platform for uninterrupted access and services. A more stable and reliable solution is necessary to ensure seamless access and maintain the trust of the user base.

10. Neurochain

Web-site: https://www.neurochain.ai/

Neurochain.ai currently does not allow users to directly rent computing power from other individuals. Instead, the platform enables users to provide their own computational resources for various tasks. For general users, the only available option at this time is to perform data tasks.

To submit a data task, users are required to fill out a detailed form where they must specify the task description, technical requirements, and provide their contact information. This ensures that the task is processed according to their specific needs.

Additionally, the platform offers a "Neurochat" feature, which allows users to interact with an AI-powered chat system based on GPT technology. This service provides users with the ability to communicate and obtain information or assistance in real-time.

10. Nosana

Web-site: https://nosana.io/

The Nosana platform requires users to complete a form and establish direct contact with the company before gaining access to computational resources. This process is mandatory for all users wishing to utilize the platform's services.

Additionally, the project does not offer the ability to view pricing information upfront. Users can only receive details about the costs of the services after submitting the form and engaging directly with the company. This approach ensures that Nosana customizes the service offerings according to the specific needs and requirements provided by the user in the form.

11. Fluence

On the Fluence platform, access to GPU computational resources is not available without first completing a form and engaging in direct contact with the company. Users must provide detailed information about their requirements through the form, after which the company will reach out to discuss the specifics and offer access to the resources.

Need to get project’s approval to get GPU

Technical side of Gintonic

1. Decentralized Infrastructure with Controller Nodes

The Gintonic platform is built upon a decentralized architecture, where a network of GPUs is managed by controller nodes. These nodes serve dual purposes:

Load Balancing: Controller nodes ensure that AI tasks are evenly distributed across the network, selecting the best available GPU for each task based on real-time performance data.

Task Hosting: They manage the lifecycle of Docker containers, which encapsulate the AI models and provide APIs for interaction.

Each controller node operates autonomously, ensuring that the platform can scale horizontally without introducing bottlenecks commonly found in centralized systems.

2. Containerized AI Models

At the core of Gintonic’s deployment model is Docker containerization. Each AI model is encapsulated within a container that includes:

• A complete API, accessible via REST endpoints.

• Hugging Face models, which are pre-trained and ready to use.

This containerization approach offers the following technical benefits:

Isolation: Each AI model runs in its own container, ensuring resource isolation and preventing conflicts between models.

Portability: Containers can be easily migrated or replicated across different GPUs in the network, facilitating dynamic scaling.

Rapid Deployment: Pre-built containers eliminate the need for extensive setup, allowing models to be deployed instantly.

3. Decentralized Controller Nodes (DCNs)

Unlike traditional centralized GPU nodes, Gintonic's Decentralized Controller Nodes (DCNs) provide a distributed, more efficient way to manage AI model workloads. These nodes intelligently route computational tasks to the nearest or most capable GPU cluster. The main technical advantages include:

Latency Reduction: By routing tasks to the nearest available GPU, DCNs reduce network latency, leading to faster task execution.

Load Distribution: DCNs dynamically balance the workload across multiple GPUs, preventing resource bottlenecks and optimizing overall network efficiency.

4. GPU Clustering for Enhanced Computation

Gintonic’s decentralized network of GPUs is organized into clusters that work in tandem to execute computationally intensive AI tasks. Each GPU cluster has the following capabilities:

Distributed Computation: Tasks are divided among several GPUs within a cluster, enabling faster completion and higher throughput.

Fault Tolerance: Clusters include redundant GPUs that can take over in case of hardware failure, ensuring continuous task execution without downtime.

Resource Pooling: Clusters share their computational resources, allowing large AI models or complex tasks to leverage multiple GPUs simultaneously.

This architecture is especially effective for scaling tasks that require heavy parallel processing, such as training deep learning models or performing real-time inference on large datasets.

Why we don’t need our own GPUs: One of Gintonic’s key advantages is leveraging an existing decentralized network of GPU clusters, eliminating the need for expensive in-house infrastructure. Instead, we rely on a reward-based model where GPU cluster providers are compensated for their resources. This approach makes our network flexible, accessible, and capable of scaling as demand for computational power increases.

5. Optimized GPU Cluster Selection with Dijkstra’s Algorithm

Gintonic uses Dijkstra’s algorithm to select the optimal GPU cluster for task execution. This algorithm assesses multiple criteria:

Performance: Identifies clusters with the highest processing power and the lowest current workload.

Availability: Ensures that the selected cluster has enough free resources to handle the task.

Proximity: Prefers clusters located geographically closer to the user to minimize latency.

By dynamically assigning tasks based on real-time data, the platform optimizes computational efficiency, reducing both processing time and resource costs.

6. Real-Time Task Execution and Billing

Gintonic’s platform supports real-time task execution, where users can interact with their deployed models through various API endpoints (e.g., for generating completions, fine-tuning models, or pre-processing data). The GPU clusters execute these tasks in real time, offering the following technical advantages:

On-Demand GPU Power: Users can tap into GPU clusters exactly when needed, paying only for the resources consumed during task execution.

Token-Based Billing: Users are billed in Gintonic tokens (GIN), which are incrementally deducted based on GPU usage, container storage, and deployment costs. This token-based system ensures transparency and real-time billing accuracy.

7. Secure Resource Management with Token Staking

To ensure that computational resources are reliably available, both controller nodes and GPU providers must stake Gintonic tokens as collateral. This staking mechanism provides several key technical guarantees:

Reliability: If a node fails to deliver computational resources as promised, a portion of its staked tokens is slashed. This incentivizes consistent performance and ensures that users receive the necessary GPU power.

Fault Tolerance: The decentralized nature of the platform means that if one node fails, another can take over, ensuring that tasks are completed without disruption.

8. Advanced API Access and AI Model Management

Gintonic offers a structured API interface for interacting with deployed AI models. To utilize GPU processing power, users must maintain a minimum balance of 1,000,000 Gintonic tokens in their accounts. Each time the user accesses GPU resources, the token balance is incrementally reduced based on the extent of GPU usage. This system ensures that users pay for GPU power in real-time, adjusting their token balance accordingly.

Within each container, the AI model is accessible via a structured API interface. An authentication key is issued to the user upon container deployment, granting access to the API, which includes the following endpoints:

• /completions/create: Generates a response based on the user's query to the AI model.

• /files/pre-process: Prepares files for fine-tuning by performing necessary preprocessing.

• /fine_tuning/create: Initiates the creation of a fine-tuned AI model.

• /fine_tuning/list: Lists all fine-tuned models associated with the user's account.

• /fine_tuning/retrieve: Retrieves the details of a specific fine-tuned model.

• /fine_tuning/cancel: Cancels an ongoing fine-tuning process.

• /fine_tuning/status: Retrieves the current status of a fine-tuning job.

• /fine_tuning/delete: Permanently deletes a fine-tuned model.

Each time a user interacts with these API endpoints, the container communicates with the decentralized GPU provisioning layer. This layer selects the nearest available GPU cluster to process the AI tasks, ensuring optimal performance and low latency.

All interactions with the API are routed through the decentralized network, ensuring that the closest available GPU handles the task, optimizing for both speed and performance.

Gintonic offers decentralized GPU network that ensures high performance and efficiency for AI tasks. It minimizes latency and maximizes computational speed. The platform's unique features include Docker containerization for easy deployment of Hugging Face models, a transparent token-based billing system, and a reward-based model for GPU providers that enhances service quality. These innovations make Gintonic a superior choice for scalable, secure, and cost-effective AI solutions compared to existing centralized and decentralized alternatives.

• Scalability - ****tasks are distributed across GPU clusters, capable of handling complex AI workloads and ensuring high performance.

• Efficiency - the use of Gintonic tokens enables real-time billing and a transparent/fair rewards system for resource providers.

• Reliability - the slashing system for GPU providers and controllers ensures that resources are delivered on time and meet quality standards.