> ## Content Index
> Fetch the complete content index at: https://aiadoption.org/llms.txt
> Use this file to discover other available public pages before exploring further.

# The fastest AI chips now reach 2.5 petaFLOP/s at 16-bit precision
- URL: https://aiadoption.org/data/ai-chip-peak-performance/
- Published: 2026-10-19T14:41:02.000Z
- Updated: 2026-10-21T06:01:17.000Z
- Description: Machine learning chips released in 2025 reach 2.5 petaFLOP/s at 16-bit precision. The best 32-bit chip is about 400 times the first chip in the data, from 2008.
- Author: Manu
- Tags: Data, AI Infrastructure, #chart, #ai-index-2026, #hub-infrastructure

2.5P

FLOP/s: peak 16-bit performance of the fastest chips released in 2025 (2.5 quadrillion operations per second)

Fastest chip by release year

20122025

## What the chart shows

Peak performance of machine learning chips has risen exponentially. The fastest chips released in 2025 reach about 2.5 petaFLOP/s, 2.5 quadrillion operations per second, using 16-bit tensor arithmetic.

At 32-bit precision the best chip in the data, released in 2022, reaches 256 teraFLOP/s, about 400 times the first chip in the series, from June 2008 (622 gigaFLOP/s).

Precision is the number of bits used to store each number. Fewer bits mean less accuracy but many more operations per second, which is why AI training has moved to 16-bit and lower formats.

## Key numbers

| Measure                      | Value           |
| ---------------------------- | --------------- |
| Best Tensor-FP16/BF16 (2025) | 2.5 petaFLOP/s  |
| Best FP16 (2025)             | 2.5 petaFLOP/s  |
| Best TF32 (2025)             | 1.25 petaFLOP/s |
| Best FP32 (2022)             | 256 teraFLOP/s  |
| First FP32 in data (2008)    | 622 gigaFLOP/s  |

## How to read it

- Peak (theoretical) performance as specified by manufacturers; real workloads achieve less.
- Tensor-FP16/BF16 uses dedicated matrix units (such as Nvidia Tensor Cores); FP16 and FP32 are standard 16- and 32-bit arithmetic; TF32 is Nvidia’s 19-bit format.
- The vertical axis is logarithmic: each gridline is ten times the one below.

**Source:** [Epoch AI](https://aiadoption.org/sources/#epoch-ai), 2026, via the AI Index Report 2026 (Stanford HAI), figure 1.2.1: *Peak computational performance of ML hardware for different precisions, 2008–25*. Chart and text by AI Adoption.  
Numbers checked against the published data on 19 October 2026.  
[Download the data (CSV)](https://charts.aiadoption.org/ai-index/csv/ai-chip-peak-performance.csv?ref=aiadoption.org) · [Cite this chart](#source)