Google announces AI chip 'TPU v5p', performance improved by up to 2.8 times from previous generation and also used for 'Gemini' training



On December 6, 2023 local time, Google announced the new model `` TPU v5p '' of the machine learning specialized processor `

`Tensor Processing Unit (TPU)' ' which is being independently developed. . Google says about this TPU v5p, ``It is also used for training Gemini , a multimodal AI that was announced at the same time as TPU v5p.''

Introducing Cloud TPU v5p and AI Hypercomputer | Google Cloud Blog
https://cloud.google.com/blog/products/ai-machine-learning/introducing-cloud-tpu-v5p-and-ai-hypercomputer



Introducing AI Hypercomputer with Cloud TPU v5p - YouTube


TPU v5p is built on 'TPU v5e' released on August 30, 2023. However, TPU v5e emphasizes power efficiency and cost performance rather than computational performance, and its actual performance was not as good as the previous model, ' TPU v4 .'

Google announces the 5th generation model of AI specialized processor TPU 'TPU v5e', up to 2 times the training performance and up to 2.5 times the inference performance per dollar compared to the previous model - GIGAZINE



On the other hand, the TPU v5p announced this time is a performance-focused TPU. Compared to the TPU v4, the TPU v5p consists of a total of 8960 chips, has 95GB of memory per chip, and has a significantly larger memory bandwidth of 2765GB per second. It has been strengthened. Google says, ``With this performance, it is possible to meet the computational needs for higher AI learning.''

Below is a table comparing the performance of TPU v4, TPU v5e, and TPU v5p published by Google. We can confirm that TPU v5p significantly exceeds the performance of conventional TPU v4 and TPU v5e in terms of the number of chips per pod, 16-bit floating point calculation performance, and the number of calculations that can be executed per second.



In addition, TPU v5p also supports learning using

int8 . If the learning speed of TPU v4 using Bf16 is 1, the learning speed of TPU v5p using Bf16 is approximately 1.9 times that of TPU v4. When used, the difference increases to approximately 2.8 times.



On the other hand, TPU v5e is better in terms of cost-effectiveness, and when training GPT3-175B, assuming the relative performance per dollar of TPU v4 is 1, TPU v5p is about 2.1 times more expensive. On the other hand, TPU v5e improves by approximately 2.3 times.



Jeff Dunn, Chief Scientist at Google DeepMind and Google Research, said, ``We saw approximately 2x speedup in training large language models using TPU v5p compared to the performance of the TPU v4 generation. 'With strong support for machine learning frameworks such as JAX, PyTorch, and TensorFlow and various automation tools, we are able to scale even more efficiently than traditional models.'

Regarding TPU v5p, Google reports, ``In fact, Google's highest-performance AI model, Gemini, which was announced at the same time as TPU v5p, was trained on TPU v5p.''

Multimodal AI ``Gemini'' with performance exceeding GPT-4, which can process text, voice, and images simultaneously and have more natural interactions than humans, will be released - GIGAZINE



Dan says, ``TPU v5p is essential for working on research and engineering using cutting-edge AI models such as multimodal AI Gemini.''

Regarding the availability of TPU v5p, Google states, ``If you would like access, please contact your Google Cloud account manager .''

Related Posts:

in Hardware, Posted by log1r_ut