Artificial Intelligence & Machine Learning

Google Cloud Next '24: AI Agents, Hypercomputers and Beyond

Google Announces Generative AI Advancements and New AI Tools to Boost Productivity
Google Cloud Next '24: AI Agents, Hypercomputers and Beyond
Image: Google

At this week's Google Cloud Next 2024 held in Las Vegas, Google announced over 1,000 product advances across Google Cloud and Google Workspace. Among the highlights were the introduction of a new AI chip, an AI agent builder, an AI Hypercomputer, a new video editing tool and enhanced chat experiences.

Generative AI was a prominent theme at the conference, and Google emphasized the seamless integration of AI tools across its cloud offerings. A key announcement was the introduction of AI agents, which could potentially transform customer service experiences.

Alphabet CEO Sundar Pichai said these advances boosted cloud revenue to a $36 billion annual revenue rate in the last quarter - "five times the run rate, five years ago."

From GPU to AI Chip Innovation

The race for cloud dominance is driven by hyperscalers utilizing powerful GPUs to process AI workloads. The surge in demand for computing power, fueled by the processing of public and private LLMs, is unprecedented.

"We are seeing computing demands of large language models grow by a factor of 10 times every year," said Amin Vahdat, vice president and general manager - machine learning, systems and cloud AI, Google.

Google, like most cloud companies, collaborates with Nvidia for GPUs; however, it aims to reduce its dependence on the GPU giant. Intel also supplies Google server-grade CPUs.

Dubbed Axion, Google announced a custom-designed AI chip to enhance the performance and efficiency of previous generations of chips used in its data centers. Axion is based on Arm architecture and will be available for preview later this year. This announcement is part of its ongoing efforts in AI/ML, complementing its custom chip designs such as tensor processing units. Customers can rent these on a pay-per-use basis.

Underscoring its capabilities, Thomas Kurian, CEO, Google Cloud, said Axion processors will deliver 50% better performance compared to x86-based virtual machines and 60% better energy efficiency than current x86 processors used in data center servers.

Google also said it will continue its partnership with Nvidia for Blackwell GB200 and NVL72 GPUs, which the company announced last month at Nvidia GTC. These will be introduced in early 2025.

Supercomputers Redefined

Tech giants have been building supercomputers for years, first with parallel processing and later, with server clusters. These involve tightly integrated tech stacks that combine hardware, software, applications, programming languages, compilers, hypervisors, runtime modules and middleware.

Supercomputers will have a new role to play in the age of AI, where processing benchmarks typically exceed teraflop levels. A teraflop is a unit of measure for calculating the speed of a computer. It is equal to one trillion floating-point operations per second. The higher the unit, the shorter the time to process extremely complex calculations - which means these computers can solve extremely complex real-world problems, such as cures for rare diseases and drug discovery.

In a blog post, Mark Lohmeyer, vice president and general manager - compute and AI infrastructure, Google, wrote, "Advancements in AI are unlocking use cases previously thought impossible. Larger and more complex AI models are enabling powerful capabilities across a full range of applications involving text, code, images, videos, voice, music and more."

Building upon these advancements, Google announced its next-generation AI supercomputing architecture with a system-level approach that combines performance-optimized hardware for storage, compute and networking. The compute infrastructure includes Cloud TPU v5p, A3 Mega VMs powered by Nvidia H100 Tensor Core GPUs, with higher performance for large-scale training with enhanced networking capabilities. It also supports open software and frameworks (Jax, TensorFlow, PyTorch/XLA) and will be offered through flexible consumption models with Dynamic Workload Scheduler.

AI Chatbot Integration

When it was first introduced, Gemini, Google's AI-powered chatbot, encountered issues with its image generation feature, sparking controversies and public protests. Google issued an apology and refined the tool, which has been undergoing testing. Pichai said Gemini 1.5 Pro has multimodal capabilities and can run one million tokens consistently. "It includes a breakthrough in long context understanding. This opens up new possibilities for enterprises to create, discover and build using AI," he said.

A token is the fundamental building block of LLM output. A single English word typically constitutes one to three tokens. Processing one million tokens indicates the LLM's capacity to handle vast amounts of information - including one hour of video, 11 hours of audio, codebases comprising over 30,000 lines of code or text exceeding 700,000 words - in one go. In a blog post, Google said it has successfully tested up to 10 million tokens. However, enterprises are concerned about the accuracy of the output, the degree of hallucinations and biases.

Google is also going a step further and integrating Gemini into its cloud offerings for software development, application life cycle, security, data analytics, business intelligence and databases.

Gemini for Workspace is an agent built into Gmail, Docs and Sheets with enterprise-grade security and privacy. With research and note taking capabilities, suggestions, data aggregation, document summarization, real-time image captioning, presentation creation and other features, Gemini is poised to streamline workflow and enhance employee productivity.

New AI enhancements in Google Meet to improve chat experiences are sure to draw long-time Zoom, WebEx and Teams users. Google announced the AI Meetings and Messaging add-on with take notes for me, chat summarization and real-time translation features. The company also announced a new AI Security add-on with privacy preserving AI models.

Google is also adding a new video assistant to its Workspace suite called Vids. Vids will integrate Docs, Sheets and Sildes to generate storyboard for video.

Gemini for Google Cloud provides AI assistance to help users work and code more efficiently, manage their applications, gain deeper data insights, and identify and resolve security threats.

AI Agents Are Here

Improving customer and employee experiences has become a top priority for enterprises, and generative AI agents can help by answering customer questions, resolving simple queries, providing customer support, and performing tasks that would take hours by humans, in a few minutes or seconds.

Google announced the Vertex AI Agent Builder which lets developers build and deploy custom models and connect them with enterprise data, systems and processes to roll out task-oriented generative AI agents.

For instance, Google is helping Walmart to create data agents to improve customer shopping experiences.

Walmart works with "massive amounts of data," said Suresh Kumar, the company's executive vice president, global CTO and CDO, and the ability to use BigQuery and other analytical tools "has been instrumental in implementing our transformation."

"Walmart is now looking at embedding AI across its businesses to unlock insights from our data to personalize experiences and get inventory to our supply chain," Kumar said.

Organizations that adopt an approach of embedding AI into all their processes to get timely and strategic insights will be better positioned to innovate and enhance their customer experiences.


About the Author

Brian Pereira

Brian Pereira

Sr. Director - Editorial, ISMG

Pereira has nearly three decades of journalism experience. He is the former editor of CHIP, InformationWeek and CISO MAG. He has also written for The Times of India and The Indian Express.




Around the Network

Our website uses cookies. Cookies enable us to provide the best experience possible and help us understand how visitors use our website. By browsing cio.inc, you agree to our use of cookies.