# Google built an AI chip with 65,536 multipliers, then asked what would make it faster: more multipliers, a faster clock, or faster memory?

- Series: [How Computers Work](https://miu.fm/howcomputerswork.md)
- Published: 2026-10-25
- Page: https://miu.fm/howcomputerswork/google-built-an-ai-chip-with-65-536-multipliers-then-asked

## Story

Google built an AI chip with 65,536 multipliers, then asked what would make it faster: more multipliers, a faster clock, or faster memory?

On a laptop we measured, one core does the same multiply-add about 31 billion times a second when the numbers are already on the chip, and about 6 billion times when each number has to come from memory first. Multiplying is quick. Getting the numbers there is slow.

In 1977 John Backus named the tube between memory and processor: "I propose to call this tube the von Neumann bottleneck." It costs energy too: on 45 nm chips, Mark Horowitz measured a fetch from memory at hundreds of times the energy of a multiply (2014).

Google found four of its six programs waiting on memory. In its own model, four times the multipliers did no better and a clock four times faster changed little; memory four times faster made the chip about three times faster (ISCA 2017).

So AI chips are now built around memory, stacked in towers right beside the chip: Nvidia's H100 (2022) moves over 3 terabytes a second, against 34 gigabytes a second for Google's 2015 chip.

The fastest chips aren't the ones that multiply fastest. They're the ones that wait least.

What would you have guessed?

Sources: our own measurement on a laptop, one core, 2026; Norman P. Jouppi et al. (Google), In-Datacenter Performance Analysis of a Tensor Processing Unit, ISCA 2017; John Backus, Can Programming Be Liberated from the von Neumann Style?, ACM Turing Award Lecture 1977 (CACM 1978); Mark Horowitz, Computing's Energy Problem (and what we can do about it), ISSCC 2014; SK hynix, 26 December 2013; JEDEC JESD235D; Nvidia, 22 March 2022 and the H100 specification page.
Sounds and music: made by us.

## Watch

- Instagram: https://www.instagram.com/reel/De6uR3zSMaW/

---

Under copyright law, a quotation must name its source: Miuradio, https://miu.fm/howcomputerswork/google-built-an-ai-chip-with-65-536-multipliers-then-asked

© 2026 Miuradio, https://miuradio.com