Aethir: A new paradigm in the field of DePIN GPU computing infrastructure

Author: Black Mario

"Aethir is the most representative and largest project in the GPU DePIN vertical track. It is reported that the project completed the TGE of the token $ATH on June 12. After it was launched to the market, the highest increase in the short term exceeded 65%. Currently, within just two days of the token entering the market, its FDV has reached more than 3.2 billion US dollars, which represents the market's high recognition of Aethir's development potential. And with the further growth of the Aethir GPU DePIN system, the ecosystem is expected to continue to create a new paradigm in the field of DePIN GPU computing infrastructure and reshape the landscape of cloud computing."

With OpenAI's successive launch of AIGC models such as ChatGPT and Sora, it is leading a new round of AI industry revolution. Although the continuous innovation of AI technology is making our production, life and working methods undergo qualitative changes, AI model training, machine learning, etc. are causing computing demand to grow exponentially. In addition to the AI ​​field, the concentrated eruption of a series of high-end industries, including cloud gaming, cloud rendering, autonomous driving, weather forecasting, and cosmic observation, also requires a large amount of computing resources.

In the traditional computing power system, mainstream cloud computing service providers usually concentrate computing power in a relatively closed manner in multiple data centers consisting of hundreds of thousands of servers, so as to continuously provide computing services to the global network. This means that the traditional computing power supply system usually has certain limitations in terms of scalability, overcoming single-point risks, and latency, and the monopoly of computing power, high construction and expansion costs are making computing resources very expensive as market demand surges. One example is that Alphago, which once defeated the Go master Lee Sedol, costs hundreds of thousands of dollars for a single training model, and commercial entities like OpenAI that need to continuously train AIGC models have to pay computing costs that can be imagined.

The rise of the DePIN narrative, which is centered on the Web3 concept, is reshaping the traditional resource allocation system. DePIN (Decentralized Physical Infrastructure Networks) is a concept proposed by the well-known encryption institution Messari. Its core is to use blockchain as a basis and use tokens to incentivize users to deploy hardware devices in a distributed manner, so as to provide a more reasonable allocation of real-world goods and services or digital resources in a lower-cost and more efficient manner. Some potential DePIN resource systems include GPU computing power, deployment hotspots, storage space, bandwidth resources, etc.

In the field of GPU computing power, Aethir is one of the most representative DePIN projects in the field of GPU computing power. By building a DePIN system with GPU computing power as the core, it is committed to addressing the conventional challenges of centralized cloud computing, including high costs, GPU supply limitations, and latency issues, and provides a decentralized GPU cloud service platform to provide long-term scalable solutions for rapidly growing markets such as AI and games, and cloud rendering.

It is reported that the Aethir network is one of the largest distributed GPU computing power ecosystems. At present, it has more than 4,000 NVIDIA H100 GPU graphics cards as computing power support, of which 82% are providing commercial services to the outside world, and it is expected that 2,000-3,000 H100 GPU graphics cards will continue to be added in the future. These computing resources will be connected to the network in a distributed manner by enterprise users, Aethir partners and individual users, which can efficiently meet the needs of the most demanding AI customers and provide enterprises with the best GPU resources worldwide.

With the continuous expansion of the Aethir distributed GPU computing power system, this ecosystem is expected to become the most important computing power cornerstone in the new technological era.

Aethir Distributed GPU Computing System

Aethir itself is a DePIN network that aggregates GPU distributed computing resources, and it runs on the Ethereum Layer2 network Arbitrum. In the Aethir network, it allows users with computing resources to connect GPU computing power to the network. The Aethir network running on the Arbitrum Layer2 network will also redistribute these computing powers in a distributed manner. Users with computing power needs can pay on demand in the form of Wholesale, Retail, etc., while users who contribute GPU resources can earn income from the Aethir network.

On the supplier side, Aethir has a wide range. Whether it is a telecommunications company, a hardware-intensive digital enterprise user, or a new infrastructure investor who holds an Aethir Edge device, as well as an individual user with idle GPU computing resources, they can all access the network and contribute to the network.

In fact, whether it is an enterprise or an individual user, GPU devices have more or less low utilization. At the same time, the ETH 2.0 upgrade is very unfavorable to the development of PoW miners. After the merger is completed, a large number of PoW hardware devices are idle (currently calculated by capital scale, the value of these idle PoW computing resources is about $19B). From the supply and demand side, on the one hand, the global computing power is in short supply, and computing resource demanders cannot afford the expensive computing costs. On the other hand, GPU idleness leads to a large amount of computing power resources being wasted. Therefore, integrating idle GPU resources will be a huge computing power resource pool, which is expected to alleviate the shortage problem faced by the computing field.

Focusing on the Aethir network itself, in order to better achieve high-quality supply and allocation of these computing powers, three main roles are introduced in the Aethir network:

• Container 

The execution role responsible for providing execution and rendering applications for the network Container, which is the actual location of virtual cloud computing. Most intuitively, we can understand it as an important driving role of computing power in the network (effective work earns income from the network).

For example, when users purchase computing power through the Internet, they use this computing power for cloud rendering, AI model training and reasoning, and even real-time response of intelligent driving. They need to use Container to provide them with computing power without delay and immediate response. By configuring high-performance hardware resources with computing and bandwidth capabilities, Container can transfer heavy-load computing tasks from user devices to the cloud, ensuring that users can enjoy high-performance computing and rendering services on various devices. Container roles are distributed all over the world, so when processing computing and rendering tasks for users, they can avoid single-point risks. To become a Container role, you need to pledge ATH tokens as the penalty cost for poor service quality, that is, for malicious acts, and it will be supervised by the Checker role.

•  Indexer

The Indexer is the resource allocator in the network. It will allocate the computing power in the network according to the actual situation of the Container (including status, service deployment, resource demand, delay and service fee, etc.) to better match the supply and demand, and realize the network's self-dynamic adjustment to ensure that the network can always maintain the best operating state. In fact, the matching of Indexer is a process of selection, that is, the better Container role is more likely to obtain the task matching of Indexer.

In the face of the demand side, the Indexer role can provide the network by setting different SDK interfaces, and can support a variety of different types of services, such as AI training, AI reasoning, cloud game real-time rendering, cloud live broadcast, etc. It can flexibly connect to the corresponding Container according to the specific needs of different users. At the same time, with the continuous expansion of the Container role in the network, Indexer can also bring full scalability to the network.

• Checker 

Checker is the supervisor role in the network. This role will monitor and check the services of the Container in the network in real time, and provide real-time feedback to the Indexer on the processing capacity, response time and network latency of the Container role, and obtain token income from the network. When a Container fails, Checker can quickly detect and notify Indexer to reallocate resources to ensure the normal operation of the network. When the service quality of the Container role processing the task does not meet the standard, Checker will lower the priority of the Container or reduce its task allocation, and confiscate it. At present, the Checker role in the network has been decentralized to the community, and there are more than 74,000 Checker roles in the network.

So overall, the Aethir network, with the mutual complementarity of the Container, Indexer and Checker roles, can not only adapt to more precise computing scenarios and ensure the SLA level of enterprise-level cloud services, but also continue to expand and innovate.

Currently, in the same track, Aethir's potential competitors include IO.net and Render Network. From the perspective of system design alone, Aethir's network architecture has excellent advantages.

One ​​is reflected in network latency. IO.net mainly relies on Mesh VPN and Reverse Tunnel technology to achieve low-latency communication between nodes. Although it has certain advantages in real-time applications, it is slightly insufficient in resource matching and service quality monitoring. Render Network focuses on off-chain rendering and on-chain payment, and there are still challenges in service real-time and quality assurance.

One ​​is in terms of customer protection. Aethir's Checker role can ensure that the network performs computing tasks uninterruptedly and continuously selects the best services, while IO.net's penalty mechanism mainly focuses on the availability of nodes and task completion, and there is relatively little real-time monitoring of service quality and prevention of fraud. Render Network can complete post-verification through Rendering Proof, but it cannot solve the problem immediately, and its real-time performance still needs to be improved.

From the perspective of ecological scale, at this stage, Aethir's cloud has $24 million worth of equipment distributed in 25 locations and 13 countries. In addition, Aethir has obtained equipment equivalent to $10 million to expand its infrastructure in 2024. Based on this system, Aethir is providing different and flexible computing power supply models based on demand.

Wholesale

For users with large and continuous computing needs, such as OpenAI, which trains AIGC models, and some large game companies, Aethir has launched a Wholesale model, which allows users to reserve the required computing processing units (PCUs) from the network in advance and obtain exclusive rendering services at a lower unit price. Demanders can make periodic expectations for computing power, such as obtaining continuous services ranging from one week to one year, and settle every 8 hours.

The advantage of this method is that it can ensure a stable and continuous supply of computing power within the cycle, and can ensure that the cost is much lower than the external market. According to the data, Aethir can provide the most favorable A100 equipment rental rate at a cost of $0.33 per hour, while Nvidia's GPU rental service, which includes 8 A100 or H100 flagship chips, costs $37,000 per month, which is equivalent to more than $6 per hour, so the Aethir network has an excellent advantage in GPU rental costs.

From the progress, Aethir has established cooperation and signed contracts with some large B-end users with annual revenue of more than $5 million, including large game studios with a player base of 150 million, WellLink, the largest cloud gaming company with more than 64 million monthly active users, and the world's largest telecommunications company, and also includes 10 other large manufacturers in the gaming field with potential cooperation.

Retail

For some users with temporary needs, such as small rendering studios, Aethir has launched a Retail model, where users can access the network at any time to obtain computing power services. The price will fluctuate according to the market price, and it will be settled in minutes.

It is worth noting that Aethir's supply market uses ATH tokens as a settlement method, which will establish a rigid demand scenario for ATH, and at the same time, through the internal circulation of the economy, it will also be able to build a more solid value foundation for the ecosystem.

Aethir Edge computing equipment

In fact, participants in the distributed computing power network are usually professional participants with certain computing capabilities, and there are certain technical barriers for ordinary users to access the network.

Recently, Aethir launched the Aethir Edge product powered by Qualcomm at the official launch of Token 2049 in Dubai. The product will serve as the only whitelisted mining DePIN device in the network, enabling global users to enjoy exclusive rewards and earn income by sharing excess bandwidth, IP addresses and computing power. By integrating local resources, Aethir Edge allows ordinary users to become nodes with excellent computing power and become beneficiaries of the sharing economy. It is reported that 23% of the total supply of $ATH tokens will be allocated to Aethir Edge holders as an exclusive incentive.

Focusing on the Aethir Edge device itself, it is equipped with a Qualcomm® Snapdragon™ 865 chipset, enabling it to efficiently handle any workload and provide continuous computing power through the Aethir network. It also has 12GB of LPDDR5 memory for processing data-intensive workloads. In addition, the device also has 256GB of UFS 3.1 storage capacity, providing high-speed data access, 1000M GE LAN network port and WIFI6 2T2R + BT5.2 connectivity.

The new launch of Aethir Edge marks a new starting point for the user-driven, decentralized innovation of edge computing.

Building the first MaaS ecosystem based on the GPU DePIN system

In fact, from the perspective of AI computing, there are also different sub-scenarios. These different scenarios have different requirements for computing power, which can be roughly divided into three categories:

• One ​​is AI training large models, which is one of the most important forms of machine training (ML) that we often talk about. Large model training usually has extremely high requirements for computing power. Only NVIDIA is unique in this field. NVIDIA H100 and A100 series chips are usually the first choice.

• One ​​is AI reasoning, which is the process of using the trained AI model to make predictions or decisions. The requirements for computing resources in this process are relatively low.

• In addition, there are some small edge-type vertical models, and this type of AI computing usually does not require such high computing power.

From the current GPU DePIN track pattern, limited by GPU resources and scale, most of the projects in this track can only meet the second and third computing requirements mentioned above.

Aethir has a clear goal, that is, to become the first DePIN project to launch Model as a Service (MaaS), deploying machine learning models to the enterprise side for users to use. Allow AI users to complete the selection and rapid deployment of open source models in one stop. Aethir MaaS will help customers achieve efficient and intelligent data analysis and decision-making, and lower the threshold for model deployment.

In order to promote the development of the ecosystem in this direction, it is building a distributed computing cluster with NVIDIA H100 GPU as the core.

Thanks to its processing unit Tensor Cores designed specifically for AI computing, thousands of parallel processing cores, high-performance bandwidth memory, AI-optimized instruction set, large-scale distributed training support, and high security and reliability, NVIDIA H100 is recognized as the most powerful GPU currently used for AI inference, machine learning, and large language model (LLM) training. At present, the Aethir network has more than 4,000 H100s through continuous expansion, and there are an additional 40,000 top-level GPUs, making it one of the largest distributed GPU computing networks at this stage.

Currently, Aethir's NVIDIA H100 GPU computing cluster is still in rapid expansion, and each NVIDIA H100 GPU added to the Aethir network needs to undergo strict inspections, including parameter configuration, model availability, bandwidth throughput, link stability detection, etc., to ensure that high-speed training and reasoning requirements can be met from a mechanism perspective. In the subsequent inspection and supervision work, it will be handed over to the third-party Checker for decentralized inspection to ensure that it provides high-quality services.

Based on its huge NVIDIA H100 GPU cluster and combined with Aethir's cloud system, it has the ability to provide MaaS services to users with large AI and ML needs for a long time, greatly reducing the technical and cost difficulties of machine learning and model training.

As the largest NVIDIA H100 GPU cluster, Aethir has formed competition with other competitors in the same track in different dimensions. It is currently the GPU DePIN project that mainly provides support for AI model training. Its potential customers are large manufacturers such as OpenAI.

Relatively speaking, io.net itself also has the ability to supply GPU computing power, but the quality and scale are far less than Aethir. Its target users are AI startups and developers, most of whom only need to perform reasoning or edge vertical model calculations, rather than AI model training.

Akash is also a potential competitor in this direction, but Akash is good at CPU network clusters. CPU cluster networks are more suitable for complex logic calculations. GPUs have more advantages in AI training, reasoning and other fields. Although Akash is also currently deploying GPU computing clusters and has also introduced H100 (only about 140), it still has a gap with Aethir in this direction. In addition to the above-mentioned projects in the same track, potential competitors such as RNDR and Gensyn are also far behind Aethir in terms of GPU computing power scale, and it is difficult to form a direct competition with Aethir in the AI ​​model training track. Therefore, scale is the advantage of Aethir's MaaS system, and driven by its own structural network, it is deeply integrated with more scenarios. On the other hand, Aethir is also establishing some potential cooperation under the new competitive landscape.

Currently, Aethir has also established cooperation with io.net, the largest GPU cluster provider. By launching 1,000 NVIDIA H100 GPUs on the io.net network, Aethir and io.net have jointly created an NVIDIA H100 GPU computing pool that is more than three times larger than its competitor Akash. This also makes Aethir the exclusive provider of io.net H100 GPUs, and continues to provide computing and MaaS services to the latter's most demanding customers who seek efficient H100 GPU processing power.

It is worth mentioning that the cooperation between Aethir and io.net is also expected to receive double rewards for H100 providers in the network, that is, while receiving Aethir rewards, they will also be able to receive double rewards of IO network tokens. This way of maximizing benefits is also an important factor in promoting the growth of Aethir H100 GPU clusters.

Reshaping the new landscape of the gaming industry

In addition to the AI ​​track, the Aethir GPU DePIN network is also of great significance to the development of the gaming field.

As mentioned above, the Aethir network has sufficient computing power on the one hand, based on which the Container role can continue to power rendering. In fact, the rendering of large-scale games requires a lot of computing resources to achieve complex ray tracing, particle effects and physical simulations. For example, the ray tracing technology solution currently widely used in the field of game rendering may require billions of ray calculations for each frame rendered, which requires parallel processing of these computationally intensive tasks. Compared with traditional computing service providers, the Aethir GPU DePIN network has more advantages in parallel processing of these computationally intensive tasks, significantly reducing rendering time and efficiency.

On the other hand, according to data from Limelight Networks, gamers generally expect the game to have a delay of no more than 100 milliseconds. Traditional service providers usually deploy game servers in data centers that are geographically close to users, and then use cloud service providers to minimize latency and optimize the player experience. But the problem is that the number of online game players in the world has reached 2.7 billion users in 2020 alone, and computing service providers need to meet the needs of these online game users at the same time, which is a big challenge for traditional cloud computing service providers. Therefore, it is particularly important to have a computing power supply system with extremely low cost, flexible scalability, and real-time, instantaneous high burst, and continuous computing power. Aethir's DePIN stack is therefore very suitable for latency-free cloud games.

At the same time, thanks to its globally distributed GPU resources that can provide efficient GPU power flow to players around the world, Aethir can also bring latency-free AAA games to low-end hardware devices in developing regions. By promoting the popularity and market share of AAA games, it is expected that the game market will continue to select and innovate, and significantly improve the quality of the game market.

In addition to support for rendering and game latency, the excellent computing power of the Aethir GPU DePIN network is also expected to help online games build better security. Focusing on the field of online games, DDoS is the most common and frequent means of attack, and all online games have paid a high cost to prevent DDoS. The Aethir GPU DePIN network can help online games resist real-time access attacks such as DDoS and ensure the continuous availability of game services.

Of course, Aethir is also committed to promoting links with the developer community, and through the launch of a series of flexible tools, including software development kits (SDKs), it can be used to integrate features such as in-game transactions and player authentication, further simplifying the gaming experience. And the analysis tools provided by the Aethir platform can help companies track player engagement, control budgets, analyze game trends, and visualize data, aiming to develop smart strategies and solve problems in the gaming industry.

Development Prospects

Aethir benefits from its network architecture design, and has certain advantages over most distributed GPU ecosystems in terms of latency, reliability, stability, and security. The Container role, which can flexibly drive computing power, has unlimited expansion capabilities, and is supervised in real time, allows Aethir to have unlimited expansion capabilities while being able to adapt to most scenarios with computing needs, rather than being directional to a specific scenario.

For example, in addition to deep adaptation to a series of scenarios such as AI, cloud rendering, and games, Aethir can flexibly adapt to scenarios with extremely high requirements for latency, including autonomous driving, and some scenarios with extremely high requirements for instantaneous parallel computing bursts, including zero knowledge and FHE. Therefore, the Aethir network itself can take root in the DePIN GPU computing track and continue to expand deeply to many scenarios with computing needs.

Aethir, which has advantages in both cost and performance, is expected to continue to seize market share from traditional cloud computing giants as its business scale and computing power continue to expand. At present, some traditional business giants with computing needs have begun to extend olive branches to Aethir, representing the recognition of these corporate users for Aethir's technical solutions and long-term development prospects.

On the other hand, as competition intensifies, the potential trend is that the Aethir ecosystem continues to welcome partners in the traditional computing field to join, forming a win-win cooperation with each other, which is extremely beneficial to promoting the exponential growth of the Aethir ecosystem and the continuous expansion of its business scale. This is not only a milestone for the Aethir ecosystem, but also a new journey in the development of the distributed GPU track.   

In fact, potential business partnerships are driving Aethir's projected annual recurring revenue of more than $20 million in Q1 2024, and an ARR of $114 million by the end of this year. As the scale of the ecosystem grows, its decentralized ecosystem will continue to form a new growth flywheel. Similarly, according to a report by Precedence Research, with the increasing application of advanced technologies such as artificial intelligence and machine learning in cloud computing, the cloud computing market is expected to exceed the $1 trillion mark by 2028, which are potential opportunities for the development of the Aethir ecosystem.

For this reason, Aethir, which has great development potential, has been able to receive support from leading Web3 investors such as Framework Ventures, Merit Circle, Hashkey, Animoca Brands, Sanctor Capital, Infinity Ventures Crypto (IVC), etc., raising more than $130 million in funds for the ecosystem and paving the way for the future of decentralized computing for Aethir.