TPU transformation: A look back at 10 years of our AI-specialized chips
Google has advanced its AI capabilities with Tensor Processing Units (TPUs), specialized chips for AI workloads, enhancing performance and efficiency, and making them available through Cloud services for external developers.
Read original article
Over the past decade, Google has significantly advanced its AI capabilities through the development of Tensor Processing Units (TPUs), specialized chips designed for AI workloads. The need for these chips arose when Google's AI compute demand began to exceed existing infrastructure, particularly for speech recognition features. The first TPU, launched in 2015, was a success, leading to the production of over 100,000 units to support various applications, including Ads and self-driving technology.
TPUs are application-specific integrated circuits (ASICs) optimized for the matrix and vector mathematics essential for AI model training and inference. The latest generation, Trillium, offers over 4.7 times the compute performance of its predecessor, TPU v5e, enabling the training of advanced AI models like Gemini 1.5.
The evolution of TPUs has paralleled advancements in machine learning, with each generation focusing on improving performance and efficiency. TPU v2 introduced a supercomputing approach, while subsequent versions incorporated innovations like liquid cooling and optical circuit switches to enhance data processing capabilities.
Google has also made TPUs available through its Cloud services, allowing external developers and companies to leverage this technology for their AI projects. Today, a significant portion of generative AI startups utilize Google Cloud's AI infrastructure, indicating the widespread impact of TPUs in the AI landscape. Looking ahead, Google plans to further customize its data center designs to optimize TPU deployment, signaling ongoing innovation in AI infrastructure.
Related
Etched Is Making the Biggest Bet in AI
Etched invests in AI with Sohu, a specialized chip for transformers, surpassing traditional models like DLRMs and CNNs. Sohu optimizes transformer models like ChatGPT, aiming to excel in AI superintelligence.
Sohu AI chip claimed to run models 20x faster and cheaper than Nvidia H100 GPUs
Etched startup introduces Sohu AI chip, specialized for transformer models, outperforming Nvidia's H100 GPUs in AI LLM inference. Sohu aims to revolutionize AI processing efficiency, potentially reshaping the industry.
What Would You Do with a 16.8M Core Graph Processing Beast?
TSMC collaborates with Intel, MIT, and AWS to develop HIVE, a graph processing unit named PIUMA with 16.8 million cores. The advanced chip aims to efficiently process neural networks, featuring a custom RISC-based architecture and photonics interconnect for scalability.
No love lost between Apple and Nvidia
Apple is using Google's TPUs for its generative AI models, diverging from the trend of Nvidia GPUs. The decision highlights Apple's focus on AI development amid ongoing tensions with Nvidia.
YC closes deal with Google for dedicated compute cluster for AI startups
Google Cloud has launched a dedicated Nvidia GPU and TPU cluster for Y Combinator startups, offering $350,000 in cloud credits and support to enhance AI development and innovation.
6 commentsTPUs are the second most widely used environment for training after Nvidia. It's the only environment that people build optimized kernels for outside CUDA.
If it was separate to Google then there a bunch of companies who would happily spend some money on a real, working NVidia alternative.
It might be profitable from day one, and it surely would gain substantial market capitalization - Alphabet shareholders should be agitating for this!
Related
Etched Is Making the Biggest Bet in AI
Etched invests in AI with Sohu, a specialized chip for transformers, surpassing traditional models like DLRMs and CNNs. Sohu optimizes transformer models like ChatGPT, aiming to excel in AI superintelligence.
Sohu AI chip claimed to run models 20x faster and cheaper than Nvidia H100 GPUs
Etched startup introduces Sohu AI chip, specialized for transformer models, outperforming Nvidia's H100 GPUs in AI LLM inference. Sohu aims to revolutionize AI processing efficiency, potentially reshaping the industry.
What Would You Do with a 16.8M Core Graph Processing Beast?
TSMC collaborates with Intel, MIT, and AWS to develop HIVE, a graph processing unit named PIUMA with 16.8 million cores. The advanced chip aims to efficiently process neural networks, featuring a custom RISC-based architecture and photonics interconnect for scalability.
No love lost between Apple and Nvidia
Apple is using Google's TPUs for its generative AI models, diverging from the trend of Nvidia GPUs. The decision highlights Apple's focus on AI development amid ongoing tensions with Nvidia.
YC closes deal with Google for dedicated compute cluster for AI startups
Google Cloud has launched a dedicated Nvidia GPU and TPU cluster for Y Combinator startups, offering $350,000 in cloud credits and support to enhance AI development and innovation.