Google's latest self-developed TPU (Tensor Processing Unit) chip is about to hit the market.
On November 6th local time, Google announced that its seventh-generation self-developed TPU chip, Ironwood, will be officially released in the coming weeks. According to reports, this chip is "tailor-made for the most demanding workloads," offering more than four times the performance in training and inference compared to the previous generation Trillium chip, making it the company's most powerful and energy-efficient custom chip to date.
The TPU (Tensor Processing Unit) is a self-developed AI-specific integrated circuit (ASIC) unveiled by Google in 2016. It is primarily used to accelerate the training and inference of machine learning models, especially deep learning tasks based on Google's open-source machine learning platform, TensorFlow. This April, Google launched its next-generation Ironwood chip for the first time.
Google states that Ironwood employs an advanced interconnect architecture to address the demands of data-intensive AI models. A single POD (AI Supercomputing Unit) can connect up to 9,216 Ironwood chips, forming a "superpod" with a breakthrough ICI (Inter-Chip Interconnect Network) capable of 9.6 Tb (terabits) per second, accessing up to 1.77 PB of shared high-bandwidth memory. (HBM), thereby eliminating data bottlenecks.
The chip has already been adopted by Anthropic, an AI startup and Google partner. Two weeks ago, Anthropic announced a partnership with Google to deploy up to one million Google TPU chips to train its large AI model, Claude. This expansion plan is worth tens of billions of dollars, with a projected computing capacity of 1 GW (gigawatt) by 2026.
James Bradbury, Head of Computing at Anthropic, said: “As demand grows exponentially, we are expanding our computing resources while pushing the boundaries of AI research and product development. Ironwood’s improvements in inference performance and training scalability will help us achieve efficient scaling while maintaining the speed and reliability our customers expect.”
Although most large language models and AI workloads still rely on NVIDIA While it uses a GPU, Google's TPU is a custom chip, which may have an advantage in cost-effectiveness for completing specific tasks. Meanwhile, in terms of performance, Ironwood represents a significant improvement over previous generations, bringing it closer to the latest chips from Nvidia and AMD.
Some analysts point out that the value of Google's TPU business may be severely underestimated. Gil Luria of research firm DA Davidson stated several months ago that Google's TPUs have narrowed the gap with Nvidia chips, becoming the latter's "best alternative." He believes that if Google's TPU business were merged with its DeepMind AI research lab, its value could reach as high as $900 billion.
Ben Reitzes, an analyst at research firm Melius Research, expressed a similar view in his report, calling Google's TPU "the most mature application-specific integrated circuit (ASIC) currently available." He stated that with the help of the TPU, Google's AI model Gemini is able to innovate rapidly, contributing to the growth of Google Cloud.
Alphabet, Google's parent company, reported third-quarter earnings last week, showing revenue of $102.346 billion, a 16% year-over-year increase; non-GAAP net income of $34.979 billion, a 33% year-over-year increase; and diluted earnings per share of $2.87. Google Cloud's growth accelerated in the third quarter, with revenue increasing by 34% year-over-year to $15.157 billion from $11.353 billion in the same period last year.
On the 6th, Google (Nasdaq: GOOGL) shares rose slightly by 0.15% to close at $284.75 per share, with a total market capitalization of $3.44 trillion.
(Source: The Paper)
