The Energy Blind Spot: NVIDIA's Flagship Edge AI Hardware Cannot Support Process-Level Energy Attribution

In the world of modern computing, a new kind of intelligence is emerging. Instead of simply answering a single question, these systems act as autonomous agents. They take a broad goal, break it down into a series of steps, call upon other tools to help, and retry if they fail. This process, known as orchestration, is moving from massive data centers to small, powerful computers sitting on desks and in vehicles. These edge devices are designed to run complex artificial intelligence locally, saving time and bandwidth. However, as these systems become more common, a critical question arises: how much energy do they actually consume? For researchers and regulators, knowing the exact energy cost of a specific task is essential. It is the only way to measure efficiency, ensure compliance with new environmental laws, and understand if these local systems are truly saving power compared to the cloud.

A team of researchers recently turned their attention to the most advanced hardware available for this job: a new generation of desktop computers built by major technology companies using a specific chip from NVIDIA. They wanted to see if these machines could tell them exactly how much energy each part of the system used while an agent was working. What they found was a significant blind spot. While the computers are incredibly fast and capable, they are completely silent when it comes to reporting their own energy usage for the central processor. The researchers discovered that the hardware has the ability to measure this energy internally, but the software that runs the machine deliberately hides this information from users.

The investigation focused on a desktop system called the ASUS Ascent GX10, which uses a powerful chip known as the GB10. This chip combines a central processor, which handles the complex planning and tool-calling of the AI agents, with a graphics processor that does the heavy lifting for visual tasks. The researchers performed a thorough audit of every possible way to ask the computer for energy data. They checked the standard interfaces that engineers use to monitor power, looking for counters that track energy usage over time, similar to a car's odometer. They found that while the graphics processor reported its power usage, the central processor offered no such data at all. There were no counters for the CPU, no power monitors on the electrical lines feeding the chip, and no standard system management tools that could reveal the numbers.

The situation is particularly striking because the data actually exists inside the machine. The chip's internal firmware, which manages the flow of electricity to keep the system cool and efficient, is already calculating the energy used by different parts of the processor. It keeps a running tally of how much energy the main processor cores use, how much the smaller efficiency cores use, and how much the graphics chip consumes. However, the manufacturers have chosen not to share this information. When the researchers asked the company behind the chip if they planned to release this data, the answer was that there were no plans to do so. This means that while the computer knows exactly how much energy it is burning to complete a task, the person using it cannot see it.

This gap matters because the way these AI agents work relies heavily on the central processor. The planning, the retries, and the coordination of tools happen on the CPU, not just the graphics chip. Previous studies have shown that this orchestration work can account for nearly half of the total energy used by an AI system. Without a way to measure the CPU's energy use, it is impossible to know the true cost of running an agent. Researchers cannot compare different software designs to see which is more efficient, and companies cannot verify if they are meeting environmental regulations. The only way to get a number today is to plug the entire computer into an external power meter, but that only shows the total for the whole machine, mixing the CPU's work with the memory, the cooling fans, and the graphics chip. It is like trying to figure out how much fuel a single engine in a car uses by only looking at the total fuel gauge for the whole vehicle.

The researchers confirmed that this is not a limitation of the physical chip itself, but a decision made in the software that controls it. They found evidence that the internal firmware is capable of exposing these numbers through a standard communication channel used by engineers, but this channel is currently turned off. They also tested a similar computer from a different manufacturer, the Acer Veriton, and found that on that specific board, the hidden data could be accessed using a community-made tool, proving that the information is there and the hardware is capable of sharing it. The difference between the two machines suggests that the missing data is a result of a product choice rather than a technical impossibility.

The paper concludes by outlining a clear path forward. The technology to fix this is already in place. The computer's internal system is ready to share the data, and the software tools needed to read it are being built by the open-source community. All that is required is for the manufacturers to update the computer's firmware to turn on the switch that allows this information to be seen. Until that happens, the most advanced edge AI computers on the market remain energy black boxes. For a field that is rapidly growing and facing increasing pressure to be sustainable, the inability to measure the energy cost of the work being done is a fundamental obstacle. The researchers argue that for these systems to be truly useful for scientific and regulatory purposes, the ability to see the energy usage must be treated as a basic requirement, just like the ability to see the temperature or the speed of the processor.