Maximum processor efficiency with lean metric

In July, a new idea was placed for a computer engineering professor at Michigan University Measuring the efficiency of a processor design, Todd Austin’s Lean metric received both praise and doubt, but even critics understood the argument: a lot of silicons are dedicated to things that are not actually computing. For example, Nvidia Blackwell GPU is more than 95 percent Nominated for other tasks, Austin told IEEE KalpanaTusrap, It is not that these parts are not doing important things, such as choosing the next instruction to execute, but Austin believes that processor architecture can move towards designs that maximize computing and reduce everything else.

What does a lean score measure?

Tod Austin: Lean stands to execute real numbers. A score of 100 percent – an acceptable target – this would mean that each transistor is calculating a number that contributes to the final results of a program. Less than 100 percent means that the design dedicates silicon and power to disable computing and logic that does not compute.

What is this other argument doing?

Austin: If you see how high-end architecture is developing, you can divide the design into two parts: the part that really calculates the program and the part that decides what to calculate. The most successful designs are squeezing that “what to do” is as low as possible.

Where is computing efficiency lost in today’s designs?

Austin: The two losses we experience in calculation are accurate losses and loss of speculation. Accurate loss means that you are using a lot of bits to calculate your calculations. You see this trend in the GPU world. They have become smaller than 16-bit from 16-bit from 32-bit floating-point proceeds. All these are trying to reduce the accurate loss in calculations.

The loss of speculation comes when it is difficult to predict the instructions. (The speculative execution occurs when the computer estimates what the instructions will happen and starts working before the instructions arrive.) Regularly, in a high-end CPU, you will see two (speculative) instruction results that are thrown for each one that is usable.

You have applied an INTEL CPU, an Nvidia GPU, and MT in RabbitAi Innerfererance Chip. Is there anything surprising?

Austin: Yes! The difference between the CPU and the GPU was very low as I thought it would be. The GPU was more than three times better than CPU. But this was only 4.64 percent (dedicated to efficient computing) vs. 1.35 percent. For Groke Chip, it was 15.24 percent. There are many of these chips that are not directly calculated.

What is wrong in computing today that you felt as if you need to come up with this metric?

Austin: I think we are really in a very good position. But this is very clear when you see AI scaling trends that we need more calculations, large access to memory, more memory bandwidth. And it comes at the end of Moore’s law, As a computer architect, if you want to create a better computer, you need to take the same 20 billion transistors and rearrange them in a way that is more valuable than the previous system. I think this means that we need a lean and lean design.

This article appears in September 2025 print issue “as” as “Tod Austin,