AI is no longer just responsive. Now, act

MIRA: Five AI lessons that the 2026 World Cup leaves us and that you can apply

Until now, most people interact with AI as if it were an assistant: ask a question and receive an answer.

AI agents work differently. Instead of just chatting, they can execute complete tasks, consult different tools, search for information in databases, coordinate several steps and deliver a result without the user having to constantly intervene.

This change implies an enormous challenge for the infrastructure.

It is no longer about answering an isolated query, but rather about keeping thousands, even millions, of agents working at the same time. To achieve this, data centers need to move increasing amounts of information between processors, memory, networks and accelerators without generating delays.

It was precisely for this scenario that AMD presented its new generation of EPYC processors.

In recent years, GPUs have become the protagonists of artificial intelligence because they are responsible for performing the most complex calculations to train and execute models. However, AMD maintains that the era of agents also returns prominence to CPUs.

While the GPUs perform the calculations, the CPUs coordinate the entire operation: managing memory, organizing data movement, and keeping the accelerators powered so they never have to stop waiting for information.

“If the GPUs are the engine, the CPU is the one who directs the traffic,” could be summarized the idea that the company tried to convey during the presentation.

With this strategy, AMD presented four versions of the EPYC 9006 family.

The most powerful, SP7, is designed for data centers dedicated to artificial intelligence and can reach up to 256 cores and 512 threads, in addition to offering 1.6 terabytes per second of memory bandwidth, a key characteristic for moving large volumes of information between the different server components.

The SP8 line targets businesses that need a balance between performance and cost, while the 9006X series is aimed at scientific simulations and high-performance computing.

Finally, the 9006 LP version was specifically designed to work alongside multiple GPUs in large artificial intelligence systems, prioritizing energy efficiency.

Beyond the technical specifications, AMD sought to demonstrate that the new platform can also reduce operating costs.

During the presentation he stated that an infrastructure made up of a thousand previous generation Intel servers could be replaced by just 82 servers equipped with EPYC Venice. According to the company, this would reduce electrical consumption by up to 77%, use 92% fewer physical servers and reduce the total cost of ownership by up to 40% over five years.

AMD also assured that the new generation multiplies by 18 the performance obtained by the first EPYC family launched in 2017 and showed comparisons against processors from Intel, Nvidia and ARM in workloads such as databases, web servers, scientific simulations and artificial intelligence inference.

For the company, this capacity will be one of the factors that will define the next stage of artificial intelligence.