Existing Problems in AI Cloud Computing
Highly Centralized
Google and Amazon are have started to provide cloud services of artificial intelligence computing. However as a commercial company, they could cut off the service at any time, given special circumstances, which are based on their own interests and the pressure of governments and other organizations. For example, Google was banned by the Chinese government, leaving Chinese users unable to use their services.
The distributed ledger function of the blockchain guarantees that it is impossible to shut down the entire network as long as more than one node is in operation. This makes it possible to design a decentralized AI cloud service which cannot be blocked.
Data Privacy Security
Although centralized companies have various security assurance agreements, it is difficult for companies to ensure data privacy when faced with internal leaks. In addition, when the government requests data, the centralized company is limited to
geographical restrictions of the host country, leaving the only option to cooperate with the government and transfer ownership of the data. As a result, the user’s data security cannot be 100% guaranteed.
The point-to-point blockchain encryption technology can be used to ensure that only the owner of the private key can access certain information, while other users cannot decrypt data nor intercept the message. This is of great significance when transmitting a variety of high-value training models and data.
Maintenance
The maintenance of centralized computing will cost heavily on man- power. The use of blockchain micro-payments makes it easy to pay for maintenance, and allows anyone to lend their computing power. The shared economy model greatly reduces maintenance staffing overhead as well as reducing computing costs.
Converting Hashpower to AI computing?
AI Computing Units (AI Mining Machine)
A lot of GPU computing mining machines can be converted into AI calculation units; this can range from simple hash calculation, to more meaningful AI task calculation.
Due to the special nature of AI computing, it is necessary to pre-install the specified system and periodically update the client, including the accounting system in order to better record hardware performance and share AI computing power.
Based on our tests, we found Nvidia Cards will be best for AI calculation. An Machine with composed of 6 *1080Ti cards which single card computing power is 7514 GFLOP/s.
A GoogLeNet model training with Caffe on 1.3 million image dataset for 30 epochs using GTX 1080Ti and Titan X video cards with 6-card parallel computing rig can be reduce the running time to 3.5 hours. But if it is using desktop it might take months.
Any mining machine that supports CUDA operations (mainly Nvidia series graphics cards) can be installed in the AI mining system. AI mining machines pre-installed common AI algorithms, such as CNN, RNN, DNN, etc., as well as a large number of other commonly used libraries, such as TensorFlow, etc. The upgraded client included the operating system, can automatically update the AI pre-installed support library.
Decentralized AI applications
The Ethereum community labels smart contract based applications as Decentralized Applications. The goal of DApp is to create a friendlier interface for the use smart contract, and to add additional features such as IPFS. DApp can run on a centralized server that interacts with Ethereum nodes. Examples include the famous Ether Delta, CryptoKitties, etc.
However, the current smart contracts are not enough for decentralized AI applications.
The reasons are the following:
• Ethereum smart contracts do not come with artificial intelligence calculations. EVM is a good contract virtual machine, but its consensus computing system can only perform simple tasks, while it is unable to perform complex artificial intelligence calculations.
• Ethereum Mining Clients do not support the computational libraries required for AI calculations. The operation of artificial intelligence depends largely on the support of various development kits, such as RNN, CNN, Tensorflow, etc.
An simple Merkle is not necessary for artificial intelligence calculation.The supporting libraries required for the relevant computing tasks can be implemented among different computing clients.
Computing Tasks
A standard payment contract will contain the following basic elements:
- AI wallet address of the task
- AI Program script of the task
- AI Output address of the task
- AI Revenue of the task
Calculation of Cost
AI calculation is usually divided into a training phase and a use phase. In the
training phase, it will use many training resources; most of the computational power will also be used here. In the use phase, due to this being the end of the training, will consume significantly less power. At the start of the mission, the Smart Contracts will pre-charge a portion of the expenses upfront. Then, the cost will be recalculated, and require the customer to pay the balance once again, to obtain the data at the end of the calculation.
In order to ensure the normal operation of the transaction, the user needs a certain amount of downpayment to begin the booking service. The automatic signing of multiple contracts will lock both sides’ funds to ensure the normal conduct of the transaction.
Sample AI applications
Creat Task-> Payment->Blockchain AI Task->AI Calculation Result
The Third Generation Blockchain — Decentralized AI Blockchain
Let everyone easily access AI computing resources from anywhere to anyplace in the world.
