New technologies are infiltrating into the machine learning market, driven by a new slate of start-up companies that are providing solutions for the conundrums pertaining to development and deployment of AI. Wave Computing is one of the leaders in the field of machine learning acceleration and is taking a fresh approach with its patented dataflow technology. Wave’s solution is different from any other on the radar screen and is widely recognized as one of the ‘hottest’ start-ups in the machine learning industry.
Wave Computing’s dataflow solution helps companies in data-driven markets of retail, finance, and more to greatly speed the creation and deployment of their large neural networks. This improves companies’ abilities to create new lines of revenue and open new market opportunities. "Wave Computing has a new kind of hardware acceleration that can do machine learning training and inferencing, faster by orders of magnitude compared to current offerings. Using our technologies, customers can now imagine things that were not possible earlier," says Derek Meyer, CEO of Wave Computing.
Wave Computing’s initial product is a 3U, rack-mountable compute appliance for the data center. It is ‘plug and play’ in a company’s existing data center configuration, and requires no new development language or APIs to be used. Initially supporting Tensorflow, Wave’s product can support other machine learning frameworks such as CNTK, Caffe, and more.
Wave Computing has a new kind of hardware acceleration that can do machine learning training and inferencing, faster by orders of magnitude compared to current offerings
Wave’s product becomes available for sale at the end of this year. But the company is now working with prospective customers to enter its Early Access Program where data scientists will be able to access a Wave prototype solution, starting this Fall.
Wave’s dataflow technology eliminates the need for CPUs, GPUs, or other co-processors such as FPGAs. The secret sauce in Wave’s dataflow technology is a custom-built ‘dataflow chip’ as well as a set of complete software programming tools and development environment. "Our solution is different. We have not attempted to replicate a CPU or co-processor. Wave's solution is a pure dataflow architecture, which directly connects a machine learning graph to our hardware, thus increasing the efficiency and scalability” adds Meyer.
Referring to one of the benchmarks Wave Computing has completed for customers, Meyer explains, Wave’s solution is able to 'train' a natural language processing application in sub-seven seconds compared to 70 to 80 minutes using existing technologies. This helps them get close to real-time training, which is one of the 'Holy Grails' of machine learning not possible today.
Looking at the roadmap, Wave Computing's vision is to extend its dataflow technology to the edge of the Cloud. Meyer added: “It’s all about the data, and putting machine learning as close as possible to the source of companies’ expanding data sets. This is part of our unique strategic plan.”