As a younger startup, SambaNova Systems is already making a mark in the fast-growing AI hardware business.
The vendor, based mostly in Palo Alto, Calif., began in 2017 with a mission of remodeling how enterprises and analysis labs with excessive compute energy wants deploy AI, and offering high-performance and high-accuracy hardware-software techniques which are nonetheless straightforward to make use of, stated Kunle Olukotun, co-founder and chief technologist.
Its know-how is being seen. SambaNova has attracted greater than $1.1 billion in enterprise financing. With a valuation of $5.1 billion, it’s considered one of the most well-funded AI startups and it’s already competing with the likes of AI chip large Nvidia.
What SambaNova gives
SambaNova’s hallmark is its Dataflow structure. Using the extensible machine studying providers platform, enterprises can specify numerous configurations, whether or not grouping kernels collectively on a single chip, or on a number of chips, in a rack or on a number of racks in the SambaNova knowledge heart.
Essentially, the vendor leases to enterprise shoppers the processing energy of its proprietary AI chips and creates machine studying fashions based mostly on area knowledge provided by the buyer, or prospects should buy SambaNova chips and run their very own AI techniques on them.
While different distributors have supplied both simply chips or simply the software program, SambaNova supplies the complete rack, which can make AI extra accessible to a wider vary of organizations, stated R “Ray” Wang, founder and principal analyst at Constellation Research.
SambaNova gives knowledge move as a service to small enterprises that lack the time, assets and need to grow to be consultants in machine studying or AI.
“The irony of AI automation is that it is massively handbook at the moment,” Wang stated. “What [SambaNova is] attempting to do is take away a lot of that handbook course of and a lot of the human error and make it a lot extra accessible to get AI.”
Wang added that SambaNova gives AI chips which are amongst the strongest on the market.
Software-defined method
While it is identified in some methods as an AI hardware specialist, SambaNova prides itself in taking a “software-defined method” to constructing its AI know-how stack.
“We did not construct some hardware considering: ‘OK, now builders exit and determine it out,'” stated Marshall Choy, vp of product at SambaNova. Instead, he stated the vendor centered on the issues of scale, efficiency, accuracy and ease of use for machine studying knowledge move computing. Then they constructed the infrastructure engine to help these wants.
The irony of AI automation is that it is massively handbook at the moment.
R ‘Ray’ WangFounder and principal analyst, Constellation Research
Two several types of prospects
SambaNova breaks up its prospects into two teams: the Fortune 50 and the “Fortune everyone else.” For the first group, SambaNova’s knowledge platform allows enterprise knowledge groups to innovate and generate new fashions, Choy stated.
The different group is made up of enterprises that lack the time, assets or need to grow to be consultants in machine studying and AI. For these organizations, SambaNova gives Dataflow as a service.
SambaNova says this method helps smaller enterprises by decreasing the complexities of shopping for and sustaining hardware infrastructure and choosing, optimizing and sustaining machine studying fashions.
This creates a “larger AI fairness and accessibility of know-how than has beforehand been held in the palms of solely the greatest, most rich tech firms,” Choy stated.
SambaNova has already attracted some big-name prospects.
One is the U.S. Department of Energy’s Argonne National Laboratory in Illinois.
Using SambaNova’s DataScale system, Argonne skilled a convolutional neural community (CNN) with photos past 50k x 50k decision. Previously, when Argonne tried to coach the CNN on GPUs, they discovered that the photos had been too massive and needed to be resized to 50% decision, based on SambaNova.
“We’re seeing new methods of computing,” Wang stated. “This method to attending to AI goes to be considered one of many. I feel different persons are going to strive totally different approaches, however this one appears very promising.”
