Donnerstag, 27. August 2026

Aube.

Nachrichten vom Fortschritt
Im EinsatzBestätigt · 2 Quellen

Waymo reveals its 1,000-TOPS chip for robotaxis

Sprachen dieses Artikels
Original · ENFR

Originaltext auf Englisch. 2 Sprachen verfügbar, Ihre kommt mit einem Klick dazu.

A passenger puts a suitcase in the trunk of a Waymo robotaxi. Beneath it sits a computer capable of up to one quadrillion operations a second, according to details Waymo has now published about the system powering its driverless fleet. The company says the hardware is already running across around 4,000 vehicles in more than 10 cities.

The most specific new component is Waymo’s custom 5-nanometer ASIC, a specialized chip that processes raw lidar, radar and camera data before it reaches the main machine-learning system. Its front-end accelerators clean up and combine those streams, run fast artificial-intelligence checks and deliver more than 1,000 TOPS — trillion operations per second — for that initial processing. Waymo says the latest system can handle data from 13 high-resolution cameras simultaneously and in real time, improving perception in low light.

Speed is only one constraint. Waymo says its compute has to withstand vibration, shocks and extreme temperatures, while remaining quiet and leaving room in the trunk for passengers’ luggage. The hardware is integrated with the vehicle’s liquid cooling system, and its architecture pairs machine-learning technology with CPUs, GPUs and other accelerators for tasks such as data movement, orchestration and logging.

The safety architecture is deliberately doubled. Waymo describes two independent compute systems running in parallel; if one experiences a fault, the other is designed to take over. The company says it has scaled its computing power 20 times in eight years, drawing on more than 200 million miles of fully autonomous experience to set its design priorities.

So what changes for riders and for Waymo? The system is already supporting approximately 500,000 paid trips a week, so the advance is less a distant laboratory demonstration than an onboard platform in service. It could help Waymo process richer sensor data without adding a human backup, while the redundant design and compact, quiet packaging protect the passenger experience. But the economics remain unsettled: estimates cited by The Verge place sixth-generation hardware at $20,000 to $25,000 per vehicle, even after a sharp fall from the estimated fifth-generation cost of $100,000 to $125,000. Waymo names AMD, Micron, Nvidia, Samsung, SanDisk, Socionext and TSMC as hardware partners, but has not disclosed its own build cost.

1,000 TOPSFront-end machine-learning performance from Waymo’s custom ASIC

Quellen — die Originale lesen(Pariser Zeit)

CleanTechnicaEN
The VergeEN
0000

Weiterlesen

Kommentare

Der Thread wird geladen…

Melden Sie sich an, um einen Kommentar zu schreiben. Anmelden