Google releases fifth-generation AI chip: accelerating the training and running speed of AI models by 5 times

Google has launched the fifth-generation custom tensor processor (TPU) chip TPU v5e for training and inference of large models. The new chip makes it five times faster to train and run AI models. Compared with the previous generation chip, TPU v5e improves training performance by 2 times per dollar and improves inference performance by 2.5 times per dollar

Google’s fifth-generation custom tensor processor (TPU) chip, known as TPU v5e, is used for training and inference of large models, resulting in training and running artificial intelligence models up to 5 times faster

At Google Cloud Next, the annual Google Cloud conference held in San Francisco, Google released a new artificial intelligence chip, the fifth-generation custom tensor processor (TPU) chip TPU v5e, for large-scale model training and reasoning. Compared with the previous generation chip, TPU v5e has improved training performance per dollar by 2 times and inference performance per dollar by 2.5 times.

Google has designed a dedicated chip TPU for neural networks, which can speed up the training and inference of machine learning models through optimization. The first-generation TPU was launched in 2016, and the fourth-generation custom processor TPU was released in 2021 and will be available to developers in 2022. Cloud TPU is a feature of Google Cloud Services and is suitable for large and complex deep learning models that require large amounts of matrix calculations, such as large language models, protein folding modeling, and drug development. Using cloud TPUs can help enterprises save money and time when implementing AI workloads

Google Cloud has launched TPU v5e, which is designed for the training and inference needs of medium and large models. This version of the chip focuses on efficiency. Compared with the previous generation TPU v4, the training performance per dollar is improved by 2 times, and the inference performance per dollar is improved by 2.5 times, while the cost is less than half of TPU v4. This enables more organizations to train and deploy larger and more complex AI models without sacrificing performance or flexibility. Google Cloud describes TPU v5e as a "supercomputer" that supports the interconnection of up to 256 chips, with a total bandwidth of more than 400 Tb/s, and offers eight different virtual machine configurations to cater for a variety of large language models and generative artificial intelligence Requirements for intelligent models. Training and running AI models is up to 5 times faster with TPU v5e, according to speed benchmarks

According to technology media TechCrunch, Mark Lohmeyer, vice president and general manager of Google Cloud Computing and Machine Learning Infrastructure, said, “This is the most cost-effective and accessible cloud TPU to date.” Lohmeyer emphasized that Google Cloud ensures that users can scale their TPU clusters to previously unreachable levels, allowing customers to easily expand their artificial intelligence models beyond the physical boundaries of a single TPU cluster. That is, a single large AI workload can span multiple physical TPU clusters and scale to tens of thousands of chips cost-effectively. "When it comes to cloud GPUs and cloud TPUs, we give customers a lot of choice and flexibility to meet the broad demand we're seeing for artificial intelligence workloads."

In addition to launching a new generation of TPU, Google Cloud also announced that it will launch an A3 series virtual machine based on Nvidia H100 GPU next month, which will be provided in the form of a GPU supercomputer to provide powerful computing power for large artificial intelligence models

The above is the detailed content of Google releases fifth-generation AI chip: accelerating the training and running speed of AI models by 5 times. For more information, please follow other related articles on the PHP Chinese website!