SAN JOSE, Calif.--(BUSINESS WIRE)--minds.ai, developers of a revolutionary scalable deep neural network training platform with dramatic acceleration performance, today announced that it has set a new record in the time taken to train AlexNet Neural Network (NN). Training AlexNet is a well-known task used for benchmarking NN training performance. The new minds.ai benchmark achieved a wall clock training time of just 2 hours and 15 minutes compared to the previous best time of 12.5 hours.
“minds.ai is proud to offer the fastest and most efficient neural network training system available today,” said Sumit Sanyal, CEO and founder of minds.ai. “This multi-server benchmark represents a major breakthrough in deep learning. We have now proven that our software can scale a single neural network training job across servers in a datacenter. Our production software will scale to support hundreds of GPUs by year’s end. Soon, every deep learning design team will be able to leverage the power of HPC to quickly develop accurate, commercial quality neural networks.”
Neural network development is usually iterative, with the results of one neural network training run informing the choices involved in designing the next iteration. Speeding up the training time of one iteration by parallelizing across multiple servers has been a major inhibitor to broad adoption and commercialization of deep learning. minds.ai expects to scale its platform up in the near future—resulting in an anticipated speedup factor of 50 times over a single GPU, reducing a month of network training time to less than one day.
Software isn’t the only factor contributing to training speed. Selection of optimal hardware is an important piece of the puzzle. minds.ai’s software is developed with specific hardware configurations in mind. The system interconnect plays a critical role in scaling to multiple servers. minds.ai’s software takes full advantage of InfiniBand’s high throughput and smart offloading in order to facilitate direct GPU to GPU communication between the servers.
The combination of minds.ai software, HPC hardware such as Quanta’s Big Sur servers, and Mellanox InfiniBand interconnect creates a scalable deep learning training solution that outperforms all other publicly available infrastructure solutions.
“We are happy to collaborate with minds.ai to deliver this performance breakthrough solution for scalable neural network training,” said Gilad Shainer, Vice President of Marketing at Mellanox Technologies. “The ability to move data faster and smarter with the use of GPU-direct technology will enable the next generation of machine learning applications.”
minds.ai (pronounced: mind’s eye) is developing a deep neural network training platform with breakthrough acceleration performance. minds.ai’s platform makes the power of High Performance Computing available to the deep learning community and equips a new generation of developers with the tools needed to quickly refine and deploy neural networks into their businesses. The company is based in Silicon Valley with offices in Bangalore, India and Amsterdam, Netherlands. More information can be found at minds.ai.