Reels

Short animated explainers on how chips work, sorted by subject. Each one plays on YouTube.

Latest reels

The newest first. New reels go up every day.

A switch with 32 optical enginesNothing on a 3 nm chip is 3 nanometresThe newest chips are printed with light from blasted tinBlow this mirror up to the size of Germany: the biggest bump is 0.1 mmDo you know the difference between all these AI chips?A 5 GB AI is 39 billion buckets of chargeYour CPU spends most of its energy not doing mathsWhy a video-game chip now trains most of the world's AIThe chip Google built because of one scary projectionWhy Amazon, Microsoft, Google and Meta draw their own AI chips

Processors and memory

How processors and memory behave, from caches to RISC-V.

Do you know the difference between all these AI chips?A 5 GB AI is 39 billion buckets of chargeYour CPU spends most of its energy not doing mathsWhy a video-game chip now trains most of the world's AIThe chip Google built because of one scary projectionWhy Amazon, Microsoft, Google and Meta draw their own AI chipsA chip that becomes a new circuit without changing a single wireWhat if one chip was as big as a dinner plate?Voice search would have doubled Google's data centersThe maths is almost free. The walk is not.This chip never goes back to memory while it worksSame chips, nearly 3x faster from software aloneSame question, different answers. Why?One AI answer uses about as much energy as 9 seconds of TVYour laptop can run an AI with the Wi-Fi offA 16 GB AI becomes 5 GB with one trickYour laptop chip is idle 99% of the timeWhat happens after you press Enter on an AI prompt?Light is slower in glass than in empty spaceNew York to London: 55 ms at the very bestA 671 GB model does not fit in eight of the best GPUsThe AI never sees your wordsThe busiest chip in the world is idle 99% of the timeOne chat can fill half a GPU: the KV cacheYou share the bus with strangers608 billion tokens in one day160 stops for every word, down to one transistor

Making chips

Silicon, light, atoms and factories.

Nothing on a 3 nm chip is 3 nanometresThe newest chips are printed with light from blasted tinBlow this mirror up to the size of Germany: the biggest bump is 0.1 mmIntel's crash, explained with a bakeryInside the transistor behind Intel's comeback chip

AI hardware, power and history

Memory, power, packaging and how we got here.

A switch with 32 optical enginesHow Marvell bought its way into lightThe race to put light inside AI chips: who is in it?Laser, shutter, waveguide: how light carries dataHow many colours of light share one glass threadZoom into an optical engine: circuits that carry lightGoogle's switch steers light with tiny mirrors3 reasons copper is still everywhere, even in AIAn AI data center is the biggest orchestra everWhat breaks first? Not the GPUAn AI data center in 18 secondsOne chip with 900,000 computers: how it worksWhy one speck of dust can kill a whole chipA giant chip full of flaws still works. How?How chips are printed with light, step by stepHow a whole wafer was stitched into one chipWhy AI chips are swapping copper for lightWhy copper wire hits a wall inside AI chipsThe signal trip that quietly eats AI powerTiny rings switch light on and off. Watch howMarvell's $115 billion bet against copperWhy one chip can never train an AI model aloneWhy Nvidia still chose copper over lightWhy optics is moving right next to the chipBroadcom's 102.4 terabit switch: why it mattersThe road from copper to light, one step at a time3 things light unlocks for AIEvery AI answer crosses this building. Here's why it mattersLight versus copper in AI chips, in 28 secondsLight is the oldest messenger. How chips use it nowTwo miles of copper inside one rack72 GPUs or 144? It depends how you count336 billion transistors in one GPUEvery GPU keeps a tower of memory beside itA switch that does the mathsOne chip stops. Everyone waits.Two giant chips act as one GPUA rack makes this much heat, so it uses water9,216 GPUs on one network: the fat treeA tray that builds in 5 minutes, not 100100,000 chips, built in 122 days8 buildings, one AI campus in TexasWhy data centers are called AI factories140 kW per rack, 1.36 MW per rowThe AI model does not fit on one GPUWhy AI racks still use copperData centers use 4.4% of US electricity

Subscribe on YouTube for new reels