Nvidia says its Vera CPU gives 1.6 times more agents for each watt

Nvidia says its new Vera CPU (the primary processor of a computer) is good for AI agents. In a test for a 100 megawatt data center, Nvidia found 1.6 times more agents for each watt than with the x86 CPU with the highest speed. All numbers are from Nvidia and its partners. The source gave no external test.

How it works

An AI agent does more than write text. It looks for data, uses databases, writes and tests code, and does tasks. A CPU does each of these steps. A data center has a limit of electricity, and because of this the work for each watt is the number to compare.

The numbers

• Against the x86 CPU with the highest speed: 1.6 times more agents for each watt.
• Against the x86 CPU with the least electricity use: 2 times the agent speed.
• Perplexity: code tasks for agents 1.5 times quicker. It will use Vera in a live system.
• Daytona: 1,000 agent sandboxes started in 10.8 seconds. With x86 it needed 88.4 seconds.

What is not known

The source gave no price. It gave no external test results. The partner tests are results that Nvidia selected to show.

Insight: Agent work makes the CPU a new fight

Nvidia is known for GPUs (processors for AI math) and at this time sells a CPU for agents. Novastart showed one computer with a maximum of 10 second agents, and AWS released a sandbox for agents. This reads as Nvidia seeking to own the processor below agent work, because each new agent adds work for the CPU. If this is correct, your agent money cost can change with the CPU, not only with the model.