Well-funded chip startup SambaNova Systems Inc. today introduced Samba-1, a generative artificial intelligence model that features more than 1 trillion parameters.

Parameters are configuration settings that determine how a neural network goes about crunching data. The more such settings an AI system possesses, the broader the range of tasks it can perform. Samba-1 is described as the third generative AI model on the market with more than 1 trillion parameters.

Palo Alto, California-based SambaNova develops machine learning chips that compete with Nvidia Corp.’s market-leading graphics cards. The company has raised about $1 billion from Intel Capital, Alphabet Inc.’s GV startup investment arm and other backers. It received a valuation of more than $5 billion following a funding round in early 2021.

SambaNova offers its silicon alongside several complementary products. There’s DataScale, a data center AI appliance that combines eight of the company’s processors with eight terabytes of memory. SambaNova also provides a bundle of machine learning software tools called SambaNova Suite that includes the new Samba-1 model it debuted today. 

“Samba-1 rivals GPT-4, however, it’s better-suited for the enterprise as it can be delivered on-premises or in private clouds so that customers can fine-tune the model with their private data without ever disclosing it into the public domain,” said SambaNova co-founder and Chief Executive Officer Rodrigo Liang (pictured).

Samba-1 is based on what the company describes as a composition of experts architecture. It comprises more than a half dozen open-source neural networks from OpenAI, Microsoft Corp., Meta Platforms Inc. and other companies. When a user enters a prompt, Samba-1 decides which of its internal neural networks is best equipped to process the request and assigns it the task of generating a response. 

The AI incorporates general-purpose large language models such as Meta’s Llama 2. Released last February, Llama 2 can generate text, craft software code and solve math problems. Samba-1 also includes more specialized neural networks such as Google’s DePlot, which can enter information from charts and other data visualizations into a spreadsheet table.

SambaNova says performing inference using the model costs 10 times less than with competing algorithms. When it receives a prompt, an AI system that comprises multiple neural networks only has to activate the one neural network selected to generate the answer. That requires less computing hardware than activating the entire system, which is necessary with more traditional AI architectures.

SambaNova customers can tailor Samba-1 for their requirements by fine-tuning its neural networks on their internal data. According to the chipmaker, a company may also set data access controls to limit which user can view what information. 

Samba-1 integrates with SambaNova’s newest AI chip, the SN40L, which debuted last September. The processor features 1,040 cores with 102 billion transistors made using five-nanometer manufacturing technology. SambaNova says that the SN40L can provide up to 638 terabytes of performance, which allows an eight-chip cluster to run AI models with as many as eight trillion parameters. 

Photo: SambaNova