This New AI Chip By SambaNova Can Run Models To 5 Trillion Parameters

Palo Alto-based AI chip startup SambaNova has unveiled its latest semiconductor marvel, the SN40L chip, poised to revolutionize high-performance computing applications. With the capability to handle models boasting a staggering 5 trillion parameters and sequence lengths of 256k and beyond on a single system node, this chip marks a significant stride in the AI hardware domain.

In a direct challenge to industry leader NVIDIA, SambaNova’s SN40L promises faster model training at a reduced total cost. NVIDIA, renowned for its GH200 chip supporting 480 GB of CPU RAM and 96 GB of GPU RAM, now faces formidable competition. SambaNova envisions the SN40L as an integral component of their SambaNova Suite, an enterprise application facilitating the creation and deployment of generative AI applications. The symbiosis between the Suite and the chip is anticipated to redefine the landscape of AI technology.

Rodrigo Liang, the CEO and co-founder of SambaNova, underscored the uniqueness of their offering. He emphasized, “We’re now able to offer these two capabilities within one chip – the ability to address more memory, with the smartest compute core – enabling organizations to capitalize on the promise of pervasive AI, with their own LLMs to rival GPT4 and beyond.”

Manufactured by Taiwan Semiconductor Manufacturing Company, the SN40L comes as a respite amidst the ongoing semiconductor shortage, especially for those seeking alternatives to NVIDIA’s chips, which are in high demand. The company’s strategic decision to integrate the chip into its custom-built tech stack, along with the SambaNova Suite, sets it apart in the competitive market. Notably, SambaNova has raised an impressive $1 billion in venture funding.

Kunle Olukotun, co-founder of SambaNova Systems, acknowledged the trend toward smaller models but emphasized the persistent allure of larger, more modular models. The company’s focus remains on delivering an LLM with the power of a trillion-parameter model akin to GPT-4, fine-tuned to individual customer data.

In an evolving landscape where competitors like NVIDIA are moving towards full-stack AI-as-a-service models, the SN40L stands as a testament to SambaNova’s commitment to pushing the boundaries of AI hardware, offering enhanced performance and superior efficiency for both training and inference, without compromising on model accuracy.

Leave a Reply

Your email address will not be published. Required fields are marked *