Behind the Compute: Building the New AI Supercomputer
As the world’s leading independent multi-modal and open generative AI company, we believe it’s important to document our journey of building Stability AI in the open as we navigate an uncharted path towards amplifying human intelligence.
Introducing "Behind the Compute," a series of blog posts that chronicle elements of our business, offering insights for others to harness the power of generative AI.
In our first post, we’ll talk about our work with Intel to build a new state-of-the-art AI supercomputer.
There are three key reasons it made sense for us to explore alternatives to GPUs when building a new AI Supercomputer to power our next-generation generative AI models:
Accessibility: In this competitive environment, quickly bringing products to market is critical. Unlike many silicon providers with lead times extending up to a year, Intel was ready to hit the ground running with their cutting-edge generative AI chip, the Intel(R) Gaudi(R)2 AI accelerator. This immediacy is invaluable in maintaining our competitive edge. Code for our most popular image and language models works seamlessly on HPUs (acronym for Intel Gaudi processor units) unlike many other accelerators that are not directly compatible.
Affordability: Simply put, the Gaudi2 offers top-tier performance at a more economical price point than its contemporaries. This affordability is crucial in facilitating rapid training and inference of our models, while managing operating costs. 8 bit precision further increases performance allowing Gaudi2 to outperform chips of the same generation.
Scalability: Intel’s software stack stands out for its seamless compatibility with our model architecture. A major advantage is the 96 GB of HBM2e memory, which is 16 GB more than competing chips, enhancing our system’s scalability to meet enterprise demands, particularly running large language and multi-modal models. Additionally, The 2.4 Gb/s interconnect is six times faster than our previous supercomputer Ezra-1, enabling large model training at almost linear scaling.
Compute power is the lifeblood of Generative AI, with a tremendous amount required to run our industry leading models. To serve our diverse community of developers and enterprise clients effectively, robust computing resources are indispensable. Often overlooked are solid alternatives to CUDA and traditional GPU solutions in technology seeking to adopt these transformative models.
We are proud to be one of the largest customers for Intel’s AI Supercomputer, which marks a significant step in meeting the relentless demand for AI computing power and offering the world a choice and an alternative.
Stay tuned for more insights in our next installment of "Behind the Compute.”
You can also stay updated on our progress by signing up for our newsletter, following us on Twitter, Instagram, LinkedIn, and joining our Discord Community.
