Google launches supercomputer architecture Hypercomputer: integrating mainstream deep learning framework and adopting flexible pricing strategy

CTOnews.com news on December 7, Google today launched a new big language model Gemini 1.0, cloud AI chip TPU v5p, but also launched the supercomputer architecture AI Hypercomputer.

CTOnews.com learned that the AI Hypercomputer computer architecture, known as a "combination of software and hardware", integrates hardware, open source software and mainstream deep learning frameworks optimized for AI, and claims to adopt a flexible pricing model that is easy for business and research departments to use.

According to Google, traditional deep learning hardware mainly relies on hardware performance to enhance AI computing speed, while AI Hypercomputer supercomputer architecture achieves a "combination of software and hardware", using the collaborative design of software and hardware to improve the efficiency of AI training.

Google claims that AI Hypercomputer supports deep learning frameworks such as JAX, TensorFlow and PyTorch, as well as software such as Multislice Training and Multihost Inferencing, and deeply integrates Google Kubernetes Engine (GKE) and Compute Engine infrastructure services.

CTOnews.com also found that Google emphasized the uniqueness of AI Hypercomputer in terms of pricing rates, claiming that the AI structure provides a more flexible consumption model, using a platform billing benchmark called "Dynamic Workload Scheduler" that provides "Flex Start" and "Calendar".

Flex Start is mainly used to fine-tune the model, carry out academic experiments and short training tasks, or carry out distillation, offline reasoning and batch tasks. The charging standard is mainly based on the amount of GPU and TPU used in AI tasks.

The Calendar mode can reserve the start time for AI tasks, which is suitable for tasks that need precise start time and model training duration. The billing standard is mainly based on the length of time.

Related reading: "training GPT3-175B model up to 180%, Google announces cloud AI chip TPU v5p"

"Google's high-profile bomb field: the official debut of the new big language model Gemini 1.0, almost completely ahead of OpenAI GPT-4."