AI startup SambaNova Techniques Inc. at this time introduced a revamped model of its flagship DataScale system, with a next-generation processor that it says vastly improves efficiency and gives assist for a lot bigger machine studying fashions than earlier than.
SambaNova is a well-funded startup that has designed and constructed an built-in {hardware} and software program platform for working AI and deep studying workloads from the information middle to the sting. The corporate says that by integrating each {hardware} and software program that’s optimized for AI, it creates a reconfigurable “information movement” structure that enables purposes to drive optimized {hardware} configurations. It implies that the underlying software program isn’t restricted by the constraints of fastened {hardware}.
The primary version of DataScale was launched in December 2020, powered by customized seven-nanometer chips. SambaNova says they’re extra in tune with machine studying and deep studying processes than the general-purpose central processing items and graphics processing items that energy most AI workloads.
Dataflow’s reconfigurable structure runs an open supply software program stack often called SambaFlow that ensures every machine studying mannequin runs optimally on the system. It does this by minimizing the necessity to work together with dynamic random entry reminiscence. That removes a key bottleneck in AI, particularly the interconnection between processors and reminiscence.
SambaNova’s DataScale platform is available for purchase or hire by way of its DataFlow-as-a-service providing. For purchasers who select the latter choice, SambaNova will set up and preserve the DataScale system of their native information middle, with pricing primarily based on utilization.
The revamped model of DataScale introduced at this time is claimed to offer numerous enhancements, together with AI mannequin coaching speeds which might be even sooner than Nvidia Corp’s broadly used DGX A100 programs. It accommodates the next-generation SambaNova processor, a enhanced Cardinal SN30 reconfigurable information stream unit. It could actually additionally assist a lot bigger AI fashions than earlier than, with 12.8 instances extra reminiscence capability than the DGX A100, SambaNova mentioned.
SambaNova mentioned the brand new system is accessible at a subscription worth solely by way of Dataflow-as-a-service. The corporate will set up and preserve the DataScale system inside the buyer’s native information middle and can cost primarily based on frequency of use.
“The brand new DataScale SN30 system achieves world document efficiency in comparison with the most recent DGX A100 programs,” mentioned Marshall Choy, senior vice chairman of merchandise for SambaNova Techniques. “With this launch, SambaNova can be providing 100% subscription pricing for DataScale and Dataflow-as-a-Service, enabling organizations to attain sooner ROI, cut back threat and scale extra cost-effectively than with the rest. different AI infrastructure.”
The corporate mentioned that corporations are more and more adopting AI to energy a variety of enterprise purposes. As such, he believes it is sensible to maneuver away from tactical AI deployments to a extra scalable, enterprise-wide resolution. That is precisely what SambaNova claims to supply with DataScale, consolidating “AI sprawl” into foundational fashions that may be skilled as soon as and reused throughout the group.
SambaNova is considerably distinctive within the AI area with its mixture of {hardware}, software program, and options, defined Andy Thurai, vice chairman and principal analyst at Constellation Analysis Inc. “For the reason that experimentation section is over, most corporations at the moment are seeking to produce scalable merchandise, enterprise-wide AI options which might be simply trainable,” he mentioned. “The SambaNova Reconfigurable Knowledge Streaming Unit chips are barely totally different than the overall function CPUs and AI-specific GPUs that the majority AI initiatives depend on, so they could be a large differentiator, and the corporate claims they’re much higher than Nvidia’s DGX A100 programs.”
One other advantage of SambaNova’s mixture of {hardware} and software program is that it gives enterprises with high-performance information processing and mannequin coaching capabilities at decrease energy consumption than competing choices, the analyst added. He defined that high-performance computing workloads have all the time been an issue for enterprises, as a result of they’re notoriously costly to coach and run.
“More often than not, corporations must buy a full suite of options to run HPC workloads,” mentioned Thurai. “With SambaNova’s subscription-based providing and higher worth/efficiency ratio, it presents a substitute for Nvidia and specialist HPC distributors. Solely time will inform if he can win.”
SambaNova Techniques is undoubtedly competing in a tricky business, however a minimum of it has a number of glad clients, together with Lawrence Livermore Nationwide Laboratory.
“We look ahead to implementing a bigger multirack system from SambaNova’s subsequent technology of DataScale programs,” mentioned Bronis de Supinski, chief know-how officer for Livermore Computing at LLNL. “Integrating this resolution with conventional clusters all through our middle will enable the know-how to have a deeper programmatic affect. We anticipate a efficiency improve of two to 6 instances, as the brand new DataScale system guarantees to considerably enhance velocity, efficiency and total productiveness.”
Photograph: SambaNova Techniques
Present your assist for our mission by becoming a member of our Dice Membership and Dice Occasion neighborhood of consultants. Be a part of the neighborhood that features Amazon Internet Providers and Amazon.com CEO Andy Jassy, Dell Applied sciences Founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and lots of extra luminaries and consultants.
– SambaNova Systems gives its integrated AI hardware and software platform a massive performance boost
