SambaNova Launches New AI Chip, Empowering Full-Stack AI Platform

Today, SambaNova Systems, headquartered in Palo Alto, launched a new artificial intelligence chip, SN40L, which will power its full suite of large language models (LLM) platform, SambaNova Suite, helping businesses build and deploy custom generative AI models.

Rodrigo Liang, co-founder and CEO of SambaNova Systems, told VentureBeat that SambaNova goes deeper into the stack than Nvidia, helping businesses train models correctly.

He explained, "Many people are very interested in our infrastructure, but the problem they encounter is that they lack expertise, so they outsource the model construction to other companies, such as OpenAI."

The "Linux" Moment for AI

Therefore, SambaNova believes this is the "Linux" moment for AI - open-source AI models will be the biggest winners. In addition to pre-trained base models, SambaNova Suite provides businesses with a curated collection of open-source generative AI models that are optimized for deployment internally or in the cloud.

Liang explained, "After we obtain the base models, we perform a series of cleaning work for businesses, including hardware optimization, because most customers don't want to deal with these. They don't want to search for GPUs or figure out the structure of GPUs."

SambaNova is not limited to chip development

However, despite not being limited to chip development and covering the entire process on the software stack, Liang insists that "compared to Nvidia, our chip has clear advantages."

According to the press release, SambaNova's SN40L can serve a model with 5 trillion parameters, and a single system node can achieve a sequence length of 256k+. It states that such design "enables higher-quality models, faster inference and training speeds, and lower total cost of ownership." Additionally, "larger memory enables true multimodal capabilities of LLM, allowing companies to easily search, analyze, and generate data for these modes."

The company also announced several additional updates regarding new models and features within SambaNova Suite:

• Llama2 variants (7B, 70B): State-of-the-art open-source language models that allow customers to adapt, scale, and run the best available LLM models and have ownership of these models.

• BLOOM 176B: The most accurate multilingual base model in the open-source community, enabling customers to address more problems across various languages and expand models to support new low-resource languages.
• A new vector retrieval-enhanced generative embedding model that allows customers to embed their documents into vector embeddings for retrieval during question-answering processes without generating illusions. Then, LLM uses these results for analysis, extraction, or summarization of information.
  
• World-leading automatic speech recognition models for transcription and analysis of speech data.
  
• Additional multimodal and long sequence length capabilities.
  
• Inference-optimized systems with 3-layer Dataflow memory for high bandwidth and capacity.