- Google’s TPU v5p AI chip is faster, has more memory, and more bandwidth than Nvidia’s H100 GPU.
- The TPU v5p is part of Google’s AI Hypercomputer architecture, designed specifically for AI applications.
- The TPU v5p is up to 2.8 times faster at training large language models than its predecessor, the TPU v4.
- Google’s TPUs are used in-house for its own products and services such as Gmail, YouTube, and Android.
- The TPU v5p is estimated to be between 3.4 and 4.8 times faster than Nvidia’s A100 GPU.
Google has released its latest version of the tensor processing unit (TPU), the TPU v5p, which is designed to power its AI Hypercomputer. The TPU v5p is Google’s most powerful custom-designed AI accelerator and offers significant improvements over its predecessor, the TPU v4. It is up to 2.8 times faster at training large language models (LLMs) and offers 2.1 times better value for money. The TPU v5p has 8,960 chips per pod, compared to 4,096 in the v4, and provides a throughput of 4,800Gbps. It also has 95GB of high-bandwidth memory (HBM) compared to 32GB in the v4.
Unlike Nvidia, which sells its GPUs to other companies, Google’s TPUs are used exclusively in-house for its own products and services such as Gmail, YouTube, and Android. The TPU v4 was estimated to be between 1.2 and 1.7 times faster than Nvidia’s A100 GPU, and the TPU v5p is estimated to be between 3.4 and 4.8 times faster than the A100, making it a formidable opponent to Nvidia’s H100 GPU. However, more detailed benchmarking is needed to draw definite conclusions.
Google’s TPU v5p marks a significant advancement in AI chip technology, as it is faster, has more memory, and offers better value for money than its predecessors. The increased speed and memory enable more efficient training of large language models, which is crucial in various AI applications. Google’s dominance in the AI chip market is likely to intensify with the introduction of the TPU v5p, as it competes directly with Nvidia’s GPUs. It will be interesting to see how Nvidia responds to this challenge and whether it can release a GPU that can outperform Google’s TPU v5p in terms of speed and performance.