[LIVE] Accelerating AI Compute with Wafer Scale
AI compute is the most important computational workload of our generation. AI has risen from obscurity to top-of-mind, with widespread and growing applications. However, it is profoundly computationally intensive. A report by OpenAI shows that compute required to train the largest models is doubling every 3.5 months, a rate 25,000 times faster than Moore’s Law.
This voracious demand for compute means that AI is constrained not by applications or ideas, but by the availability of compute. Testing a single new hypothesis — training a new model — takes weeks or months and can cost hundreds of thousands of dollars in compute time. This is a significant drag on the pace of innovation. Google, Facebook, and others have noted that long training time is the key impediment to progress in AI - many great ideas are ignored because they take too long to train.
Dhiraj Mallick, VP of Engineering and Business Development at Cerebras Systems, will discuss how AI’s true potential can be realized through eliminating this primary impediment to the advancement of AI — by reducing the time it takes to train models from months to minutes, from weeks to seconds. Cerebras recently introduced CS-1, which is comprised of their Wafer Scale Engine (WSE), the first and only trillion transistor processor for AI applications, the CS-1 system, which delivers power, cooling and data to the WSE, and the Cerebras software platform that enables quick deployment of the full system and allows researchers to use their existing software models without modification.
In this session, Mallick will share perspective on the state of the AI industry and technologies that will enable AI to continue to rapidly grow and evolve.