Hardik's Notes · AI, simply

Five million times faster? Read the small print

Arduino compared a 2000s chip with its newest AI board. The headline number is huge. The way they got it shows how to read any speed claim.

A small chip rides a skateboard beside a measuring tape. A big glowing chip sits on a rocket. Two piles of coins in different colours lie in front.

What is it?

On 20 October 2026, Arduino’s chief product officer, Marcello Majonchi, wrote a post comparing two boards. One is the Arduino UNO Rev3, with a chip from the 2000s. The other is the VENTUNO Q, Arduino’s most capable AI board so far.

The VENTUNO Q is listed at 40 “dense TOPS”. That means forty trillion operations a second. Arduino estimates the old chip at about 8 million. Divide one by the other and you get five million.

Why is that number not fair?

Chips count in different “currencies”. Small whole numbers need little work. Numbers with decimals need more. AI chip makers usually quote the small kind. The old chip’s estimate counts general work. Comparing the two is like comparing rupees with dollars without an exchange rate.

What is a fair race?

Give both chips the same job. Arduino picked the maths at the heart of AI: multiply two numbers, add the result to a running total. It is called multiply-accumulate. Image recognition is millions of these.

The job: look at one photo, 224 by 224 pixels, and decide what it shows. That is about 8 billion operations.

17 min

Arduino's estimate for the old UNO to check one photo, against about 5,000 photos every second on the VENTUNO Q

Arduino blog, 20 October 2026

You can check this yourself. 8 billion operations at 8 million a second is 1,000 seconds. That is about 17 minutes. For 5,000 photos it would take about 58 days.

What’s the catch?

The old chip cannot run this job. The AI model needs roughly 100 megabytes of memory. The old chip has 2 kilobytes of memory and 32 kilobytes of storage. The 58 days only works if you imagine unlimited extra memory.

Arduino says this itself. Its point: big jumps are not “the same thing, faster”. They are new things becoming possible. No amount of patience fits a 100 megabyte model into 32 kilobytes.

Also remember: this is one company’s estimate on its own boards. No outside paper backs it.

Try it

Pick any “10 times faster” claim you have seen. Write down three questions. What was measured? Was it the same job on both? Could the older one do it at all? Then do the division yourself, as above.

Sources

Beyond the notebook

This is what we do in schools.

A year-long programme for classes 6 to 12. Creativity, taught as a skill, inside the classroom you already have, led by your teachers.

See the programme