Highlights:

  • SambaNova has secured approximately a billion USD in funding from Intel Capital, Alphabet Inc.’s GV startup investment arm, and other investors.
  • SambaNova customers can customize Samba-1 for their specific needs by fine-tuning its neural networks using their internal data.

Well-capitalized chip startup SambaNova Systems Inc. introduced Samba-1, a generative artificial intelligence model boasting over 1 trillion parameters.

Parameters are configuration settings that dictate how a neural network processes data. The greater the number of such settings that an AI system has, the more tasks it can execute. Samba-1 is positioned as the third generative AI model available, boasting an extensive parameter count exceeding one trillion.

SambaNova, a Palo Alto, California-based company, creates machine learning chips that compete with the industry-leading graphics cards from Nvidia Corp. The company has secured approximately a billion USD in funding from Intel Capital, Alphabet Inc.’s GV startup investment arm, and other investors. It achieved a valuation of over USD 5 billion following a funding round in early 2021.

SambaNova offers its silicon alongside several complementary products. Among these products is DataScale, an AI data center appliance that combines eight processors from the company with eight terabytes of memory. Additionally, SambaNova provides the SambaNova Suite, a collection of machine learning software tools that includes the recently released Samba 1 generative AI model.

Rodrigo Liang, Chief Executive Officer and Co-founder of SambaNova, said, “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.”

According to the company, Samba-1 is built on a composition of expert architecture. It includes more than six open-source neural networks from companies like Microsoft Corp., OpenAI, Meta Platforms Inc., and others. When a user enters a prompt, the Samba-1 generative AI model determines which of its internal neural networks is best suited 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 into its architecture. Top of Form

Released last February, Llama 2 can generate text, crafting software code, and solving math problems. Samba-1 generative AI model also incorporates more specialized neural networks like Google’s DePlot, which is capable of extracting information from charts and other data visualizations into a spreadsheet table.

According to SambaNova, executing inference with the model is ten times less expensive than utilizing rival algorithms. When it receives a prompt, an AI system comprising multiple neural networks only needs to activate the selected neural network to generate the answer. This requires less computing hardware than activating the entire system, as is necessary with more traditional AI architectures.

SambaNova customers can customize Samba-1 for their specific needs by fine-tuning its neural networks using their internal data. According to the chipmaker, a company can also establish data access controls to restrict which users can view specific information.

Samba-1 integrates seamlessly with SambaNova’s latest AI chip, the SN40L, which made its debut last September. The processor boasts 1,040 cores with 102 billion transistors, all fabricated using five-nanometer manufacturing technology. SambaNova states that the SN40L can deliver up to 638 teraflops of performance, enabling an eight-chip cluster to execute AI models with as many as eight trillion parameters.