The Infrastructure Agnostic Platform

On the backdrop of a Ethereum’s massive fall in 2018, we have received many questions about the future of Ethereum mining facilities and the use of Ethereum mining rigs for knowledge mining on the Neuromation Platform.

We addressed this in our Q2 Report. Here we offer a more complete explanation of the testing, data, conclusions and actions we have taken re: computing power and Ethereum mining rigs on the Neuromation Platform.

TL;DR

In regards to computing power on the Platform, Neuromation began with the hypothesis that distributed computing power can provide vast, low-cost computing resources for the training of machine learning (ML) algorithms. Specifically, Neuromation sought to integrate Ethereum mining rigs for ML training to accelerate for the adoption of Artificial Intelligence (AI) globally.

Neuromation continues to believe that distributed computing power holds promise for ML training. But, at this point, market, technical and security demands make it clear that the most prudent approach to our goal of AI adoption through the Neuromation Platform is computational agnosticism. Therefore, Neuromation will continue to build the Platform for integration into computational platforms that comply with market demands. Currently, those demands are for large cloud-based platforms. We will review this policy as market technical and security demands change.

DETAIL

I. Neuromation original hypothesis

• Global distributed computing power, including Ethereum mining rigs, can provide vast computing capacity for the training of ML algorithms on large datasets
• The use of mining rigs for ML represents material cost savings as compared to popular cloud-based computing services provided by the major tech companies, helping to driving widespread adoption of applied AI

II. Testing Completed

Market Demand

• During 2018, Neuromation engaged potential clients across global AI markets in multiple countries in Asia, Europe, Australia and North America
• It became clear to us that a vast opportunity for our AI-related products and services comes from enterprise clients
• We believe that, as global AI adoption grows, the profiles of potential clients for AI will expand — but, at the current time, a focus on enterprise clients is the most sensible strategy for the development of the Neuromation Platform
• There are at least two key elements that are fundamental to contracts with enterprise clients. The first: the requirement that service providers adhere to data storage and security policies. The second: the requirement that technology providers provide stable and predictable services that meet client expectations
• In addition to the above, many enterprise clients maintain existing contracts with large-scale cloud-based computing providers such as AWS, Google Cloud, Azure, others

Technical

• During the 1st and 2nd quarters of 2018, Neuromation tested the use of GPU Ethereum mining rigs for use in deep learning applications at scale
• Our Engineering and Research Team worked with standard benchmarks in object detection for well-known datasets
• While NM’s testing was specifically related to object detection (computer vision), we believe that the results of the testing can be abstracted to AI training tasks across computer vision and other domains
• These tests were duplicated on AWS servers (as an upper bound) and comparisons were made
• The results of our testing clarified two major technical problems that need be solved in order for knowledge mining to be adopted at scale. These problems are as follows:

THE DATA SECURITY PROBLEM

• Ethereum miners do not typically have in place the level of data security required by enterprise clients
• The costs of capital and time required to implement these security minima are prohibitive

THE NODE INTERCONNECTION PROBLEM

• The architecture of tested Ethereum mining rigs is not optimized to provide computational interconnectivity that is a factor in the training of deep learning algorithms, resulting lower efficiencies as compared to cloud-based compute provider
• The solution to these problems may require: a) increased data security at mining rigs: the development of data security technology (we believe that blockchain provides one potential avenue for development) that will reduce the need for on-site data security systems, b) changes in client profile as the AI sector develops, c) modifications to mining infrastructure as new mining hardware comes on line

Business

• Over 2018, the business environment for Ethereum miners deteriorated dramatically, increasing the likelihood of insolvency of mining facilities, with specific instances of bankruptcy
• While these changes may eventually be net positive, Neuromation believes that stability must come to the mining sector prior to undertaking scaled use for contract clients

Economic

• As compared to AWS cloud based servers, Ethereum miners provided a material net savings for the training of AI algorithms (AWS was 2–3x times higher than mining rigs), supporting the Neuromation’s hypothesis on the cost savings delivered by distributed computing — but this economic advantage does not outweigh limitations demanded by the market

III. Conclusions

• Neuromation continues to believe that distributed computing power holds promise for the training of deep learning models
• Ethereum miners provide material cost savings to cloud based providers
• Ethereum miners have multiple problems that need to be solved. These include: a) data security, b) miner node interconnectivity, c) stability of miners’ business models, d) net competitive disadvantage over existing cloud based services
• Neuromation sees the bulk of current demand for AI products and services coming from enterprise clients, warranting a focus on these clients as early adopters of the Platform
• Neuromation believes that the “Computational Agnostic” policy continues to be the most reasonable position for the Platform. We believe that this policy will drive faster adoption from our current and future clients, allow us the greatest flexibility as the market evolves, and adhere to the most responsible use of resources to ensure success of the Platform over the long term
• Neuromation will continue to monitor developments in the mining community and seek to adopt viable solutions across the spectrum of distributed computing power, including the scaled use of Ethereum mining rigs, as their technical and economic limitations are resolved