AI trend is on the rise again: DePIN protocol for computing power aggregation is disassembled in depth io.net

Author: Climber, Golden Finance

Recently, AI has once again become the focus of the market. OpenAI first launched GPT-4o, and then Google released dozens of Google and AI combined products at the I/O Developer Conference. The AI ​​business competition is intensifying.

At the same time, io.net, a popular project in the AI+Web3 track, is also actively preparing for the battle. As a project that superimposes the concepts of AI + DePIN + Solana, io.net is currently carrying out the first season of Ignition rewards activities, and its token IO contract address has been announced. Bybit Pre-Market has also launched futures trading of io.net (IO). It is believed that IO spot will also be launched on major CEXs soon.

The biggest feature of io.net is that it can aggregate users' idle GPU computing power to provide cloud computing services for users with AI needs. Under the hot trend of AI, Golden Finance will fully interpret the io.net network that is benchmarked against AWS (Amazon Web Services) and deeply understand its principles, mechanisms and various key information.

io.net Introduction

io.net is a decentralized AI computing platform based on Solana developed by IO Research. The project mainly aggregates GPU resources for the entire life cycle of AI and ML company workloads (from model training to inference), and is committed to providing low-cost and efficient cloud computing services.

io.net can aggregate multiple GPUs from independent data centers, cryptocurrency miners, and crypto projects such as Filecoin or Render to obtain computing power, thereby providing cloud computing services to target companies with more cost-effective advantages.

So why does io.net aim at GPU resource integration?

Before figuring out this question, we must first understand what GPU is. According to Intel's official website, a graphics processing unit (Graphics Processing Unit) is a microprocessor composed of many chips for image computing work.

At the same time, according to AWS, one of the commercial application scenarios of GPU is blockchain. Since cryptocurrency is built on blockchain, special types of blockchains (proof of work POW) usually rely heavily on GPUs for operation.

This shows that GPU has a broad market demand, but currently there is a huge gap in this business demand.

Currently the major cloud providers have about 10-15 exaFLOPS of available GPU computing power, but due to the surge in demand for AI/ML model training and inference workloads, the potential demand for cloud GPU computing may be as high as 20-25 exaFLOPS. The difference between the two is in the range of 5-10 exaFLOPS, which means that cloud GPU capacity needs to be expanded 2-3 times from the current level to fully meet user needs.

The development cycle of GPU supply is long, so io.net solves this problem by accessing underutilized GPU resources outside the cloud, such as independent data centers, cryptocurrency miners, consumer-grade GPUs, etc.

Product composition

io.net's architecture includes the underlying IO network, IO engine, and IO Element. The IO network is a decentralized GPU network that uses a mesh VPN architecture to provide ultra-low latency communication between nodes and ensure high efficiency of GPU resource sharing and processing. The IO engine is a programmable computing layer built on the IO network. Together, these elements form a complete ecosystem dedicated to providing AI computing solutions.

io.net connects users and computing power providers through three main products:

IO Cloud: deploys, manages and allocates decentralized GPU clusters on demand, and provides a comprehensive solution for scaling artificial intelligence and Python applications through seamless integration with IO-SDK.

IO Worker: provides users with real-time insights into their calculations, provides operations and a bird's-eye view of devices connected to the network, allowing them to monitor these devices and perform quick operations such as deleting and renaming devices.

IO Explorer: mainly provides users with comprehensive statistics and visualizations of various aspects of the GPU cloud, allowing users to easily and instantly monitor, analyze and understand the complex details of the io.net network, providing full visibility of network activities, important statistics, data points and reward transactions.

Core Features

Batch Inference and Model Serving: Inference can be performed on incoming batch data in parallel by exporting the architecture and weights of the trained model to shared object storage. io.net allows machine learning teams to build inference and model serving workflows across distributed GPU networks.

Parallel training: CPU/GPU memory limitations and sequential processing workflows present a huge bottleneck when training models on a single device. io.net leverages distributed computing libraries to orchestrate and batch training jobs so that they can be parallelized across many distributed devices using data and model parallelism.

Parallel hyperparameter tuning: Hyperparameter tuning experiments are inherently parallel, and io.net leverages distributed computing libraries with advanced hyperparameter tuning to check for optimal results, optimize scheduling, and simply specify search patterns.

Reinforcement learning: io.net uses an open source reinforcement learning library that supports production-level, highly distributed RL workloads and a set of simple APIs.

Key data

According to the io.net official website, the project provides computing services at 90% cheaper prices than traditional cloud providers. In addition, io.net enables users to access provisioning and deploy clusters within 90 seconds to obtain GPU computing power from the cloud. Therefore, the combination of speed and cost makes io.net 10 to 20 times more efficient than traditional cloud products.

In April this year, io.net announced that the network has more than 20,000 A100 GPUs and more than 520,000 GPUs and CPUs, and the total value of io.net's infrastructure has exceeded the $2 billion mark.

In terms of team size, io.net has more than 100 people.

According to other data, within four months of the project's birth, io.net has provided more than 50,000 working nodes and more than 60,000 computing hours.

The current total profit of the io.net network is about 1.02 million US dollars, the total computing time of the service is 132,000 hours, and the number of GPU clusters created exceeds 1,100.

Financing and Economic Model

In March this year, io.net completed a US$30 million Series A financing, led by Hack VC, with participation from Solana Labs, ArkStream, Animoca Brands, OKX Ventures, and investors including Solana founder, Aptos founder, and Animoca Brands founder.

In April this year, io.net announced the token economic model of the IO token. The fixed maximum supply of the IOG network's native token IO is 800 million, of which 500 million IO tokens will be distributed at the time of release, including seed round investors (12.5%), A round investors (10.2%), core contributors (11.3%), R&D and ecosystem (16%) and community (50%).

With the issuance of IO to incentivize network growth and adoption, it will grow to a fixed maximum supply of 800 million in 20 years. Rewards will be released to suppliers and pledgers every hour for 20 years.

The rewards adopt a deflationary model, starting from 8% in the first year, and decreasing by 1.02% per month (about 12% per year) until the upper limit of 800 million IO is reached. As rewards are issued, the share of early supporters and core contributors will continue to decrease.

Main Actions

To incentivize GPU computing power providers, io.net launched two seasons of Ignition rewards, each offering 7.5 million tokens.

On May 14, io.net integrated Aethir's more than 300 NVIDIA H100 GPUs, and the two are expected to deploy a total of 1,000 H100 GPUs by the end of May.

On May 9, io.net announced the launch of Ignition Season 1 points. On May 1, io.net added a GPU Proof of Work mechanism. On April 27, the io.net team prevented the retaliatory attacks of banned airdrop hunters, restored all device data, and launched multiple security patches.

In mid-April, io.net reached cooperation with Aethir, Mind Network, and Bitcoin L2 project BVM.

Summary

Io.net said its goal is to build the "GPU Internet" and create the world's largest artificial intelligence computing DePIN to solve the shortage of GPU computing power caused by the artificial intelligence boom. From this point of view, the project has a bright future and a strong financing background.

However, as an emerging project, it also faces technical security risks, such as hacker attacks before the mainnet went online. In addition, the business competition is fierce, facing challenges from AWS, Google Cloud, Alicloud and other companies. Therefore, investors do not need to be too enthusiastic about Io.net.