Customers of Google Cloud can now use the sixth generation of Trillium TPUs. This chip, also known as v6e, is used for training and deploying artificial intelligence. Google promises higher performance, bandwidth, and energy efficiency compared to the previous chip generation. The company announced this chip generation in the spring and has been using it to develop the second iteration of its in-house AI model, Gemini, which is currently available as an experimental version.
Trillium offers four times the computing power. Like the v5e chip, customers can train, fine-tune, and deploy transformers, text-to-image generators, and convolutional neural networks with the Trillium TPUs. Compared to its predecessor, the Trillium TPU’s performance data allows for four times the training performance of AI models and triple the inference throughput. Google also promotes the v6e accelerators with 67% higher energy efficiency.
Google specifies the computing power for 16-bit floating-point numbers for Trillium accelerators at 918 teraflops, while v5e chips reach 197 teraflops. A v6e pod with 256 chips achieves about 235 petaflops. The values for 8-bit integer calculations are twice as high. With a capacity of 32 gigabytes and a bandwidth of 1.6 terabits per second, Google doubled the values of the high bandwidth memory. Each Trillium chip features a tensor core with two units for matrix multiplication (MXU) and one each of vector and scalar units. The tensor cores of the v5e chip contain four MXUs.
Trillium instances are twice as expensive as their predecessors. In Europe, Google offers Trillium instances at the Amsterdam data center. On-demand, a chip hour costs $2.97. With a one-year commitment, the price is $2.08, and for three years, it is $1.34. Instances with the previous v5e cost about half as much. For spot VMs, dynamic prices apply. Other locations for Trillium instances currently include South Carolina, Ohio, and Tokyo.
Google’s introduction of the v6e chip marks a significant advancement in AI processing capabilities. The increased performance and efficiency are crucial for handling the growing demands of AI applications. This new generation of TPUs will likely enable more complex and faster AI model training, which is essential as AI continues to evolve and expand into more areas.
The Trillium TPUs’ enhanced capabilities are expected to benefit a wide range of industries and applications. From natural language processing to image recognition and beyond, the improved processing power and efficiency can lead to more accurate and efficient AI models. This is particularly important in fields that require real-time data processing and analysis, such as autonomous vehicles, healthcare diagnostics, and financial services.
As AI models become increasingly sophisticated, the need for powerful computing resources like the Trillium TPUs becomes more apparent. Google’s investment in developing these advanced chips demonstrates its commitment to pushing the boundaries of AI technology. The availability of these TPUs on the Google Cloud platform provides businesses and researchers with the tools needed to innovate and drive progress in AI.
Overall, the launch of the v6e Trillium TPUs is a significant milestone in the AI industry. It highlights the rapid advancements in AI hardware and the ongoing efforts to enhance AI capabilities. As more organizations adopt AI technologies, the demand for high-performance computing solutions will continue to grow, making innovations like the Trillium TPUs essential for the future of AI.