Back in April, SambaNova announced that it had raised $676 million through a Series D round led by SoftBank Vision Fund 2, the Japanese holding company’s investment fund. The funding was used, among other things, to develop its innovative “Dataflow as a Service” platform, a combination of AI-based hardware known as Datascale and the Sambasystem that runs it. With this solution, the start-up wants to bridge the gap between companies and artificial intelligence.

The DataScale tool and the SambaNovaFlow

system from the start-up SambaNova
The startup SambaNova was founded in 2017 by Rodrigo Liang, a former Oracle employee, with the support of two Stanford University professors Kunle Olukotun and Chris Ré. The former contributed to the discovery of the multicore processor and the latter was awarded the MacArthur Fellowship, which allows awardees to pursue and develop their business with the support of the MacArthur Foundation.

The start-up is known for having developed a platform dedicated to AI called DataScale SN10-8R (abbreviated to DataScale) which is based on an AMD Epyc x86 server with 12TB of DDR4-3200 and equipped with eight Cardinal SN10 custom accelerators (see illustration) linked together via a dedicated switch. A Linux distribution associated with SambaNovaFlow software is used to run this base. DataScale is thus an integrated platform of software and hardware systems.

The chips that make up the platform concentrate reconfigurable units (RDU) optimizing the flow of data in order to accelerate the computing performance and ensure an optimal bandwidth. According to the firm, the circuit can scale if a workload needs more memory with much better performance and efficiency than a structured FPGA/ASIC. Below is a schematic explaining the RDU data flow in DataScale:

“Dataflow-as-a-Service”

Last December, the startup launched “Dataflow-as-a-Serivce”, its subscription-based service that allows companies to leverage the AI system designed by SambaNove, and do so on demand. Rodrigo Liang, the firm’s founder and CEO, said around the creation of the service:

“Rather than coming in and thinking about how to hire scientists and then deploy an artificial intelligence service, you can now subscribe and bring this technology overnight. We’re very proud that our technology is pushing the boundaries of how artificial intelligence can be used in the industry.”

So this service is based on the premise that companies want to integrate more AI into their operations or development strategies, but may sometimes lack the skilled staff or money to get it right. Companies focus on their AI model without worrying about the technical part, which is provided by the US startup through DataScape and SambaNovaFlow. Below, find an illustration that shows SambaNova’s “Dataflow-as-a-Service” service:

Helping Companies in Industry 4.0

SambaNova claims that its tool will deliver higher performance than GPU-CPU configurations for TensorFlow or PyTorch-based algorithm models. As shown in the diagram below, the main machine learning tools are supported by the firm’s platforms and systems:

DataScale is one of the many possibilities that companies can exploit to get a little more into the world of AI. Other companies, start-ups as well as giants in the technology sector, are trying to offer innovations in edge computing, cloud computing or computing power: Alibaba with Xuantie 910, Google which has announced its fourth generation of its TPUv4 chip or Nvidia which wants to launch its new ARM Grace processor for 2023.

More and more companies seem to be interested in processors specifically designed for them or for applications in industry that can efficiently run AI or machine learning models. With this craze, the number of new technologies could be multiplied in the next few years.

Translated from Comment SambaNova propose des outils permettant aux entreprises d’exploiter l’intelligence artificielle ?