Bitmain Introduces Its First Hardware for Accelerating Artificial Intelligence (AI) Applications
by Sergey Nikolenko, Chief Research Officer Neuromation.io
Yesterday, Bitmain announced the release of a new customized ASIC for tensor computing, that is, primarily for training deep neural networks. Previously, basically the only custom hardware for machine learning have been Google TPUs, which have never been released for the general public and probably will not be released in the near future. It is great to hear about competition in this niche.
Here at Neuromation, we are developing a platform whose basic idea is to tap into the huge computational resources that are currently spent on mining cryptocurrencies. There are millions of GPUs in the world that have been doing basically completely pointless work, solving harder and harder computational problems that have no real world relevance other than generating a little bit of money for the owners.
With our platform for “knowledge mining”, the miners will be able to pool their resources to empower synthetic data generation (another strong suit of Neuromation), training of large machine learning models, and model deployment for production purposes — all the while getting much more money than by mining ETH or other cryptocurrencies. However, it is hard to repurpose special hardware such as ASICs designed for bitcoin mining to deep learning — special hardware usually supports only very specific algorithms. Thus, so far we have envisioned a GPU-based platform.
This exciting news may bring completely new possibilities to the distributed AI platform that we are developing at Neuromation. Similar to how large BTC mining pools have arisen based on specialized ASICs for BTC mining, we can now envision large training pools of specialized hardware for training machine learning models — smaller perhaps than Amazon Web Services and Google Cloud, but similar in design and intent. And competition will naturally drive the prices down, which is always a good thing and lines up perfectly with our goals of democratizing AI for everyone.
But the important thing is with AI model training there is no such thing as simply buying hardware and putting it to work, like the miners have been doing with cryptocurrencies. You also need to have the problems, the datasets, — the clients who are willing to pay for model training.
That is why the new “knowledge miners”, even armed with special purpose hardware, will require a distributed platform that will bring them together with their clients: AI startups, individual AI practitioners, in the future perhaps even automated agents that can outsource parts of their training to external computational devices. This platform is exacly our vision here at Neuromation, and that is why we are very excited about the news from Bitmain.
