特斯拉削减AI5和AI6芯片内存以保障Optimus机器人量产

Tesla is trimming the memory on its next-generation AI5 and AI6 chips, and Elon Musk says it comes down to getting enough of them built. The cuts are meant to help Tesla scale AI5 and AI6 production, particularly for the Optimus humanoid robot.

What Musk Said About AI5 and AI6 RAM

Musk shared news of the change on X earlier this week, saying Tesla cut RAM in half for AI5, to 72GB of LPDDR5, and by a third for AI6, to 144GB of LPDDR6. He called it “the only way to get enough volume for Optimus production,” and said the move also greatly reduces cost.

As for performance, Musk said Tesla expects “a negligible effect on Optimus performance.” That’s due to memory bandwidth limiting these systems more than total capacity does, and overall bandwidth hasn’t changed.

He later made a small adjustment to the plans. AI5 is being bumped up slightly to 96GB, since “Tesla would otherwise be the only company using the min RAM version of LP5.” So AI5 ends up with a third less memory than originally planned rather than half, much like AI6.

How That Compares to Current Hardware

The numbers are still a major step up compared to what’s in Tesla’s cars today. AI4 computers ship with 32GB of GDDR6, made up of 16GB per chip across two chips on each mainboard. Hardware 3, in comparison, had just 8GB of LPDDR4 RAM per chip, but bandwidth was the real bottleneck there, at roughly 15% of AI4.

Tesla is also working on a mid-generation AI4+ (or AI4.5) computer upgrade with more memory and compute, expected to double AI4’s memory to 32GB per chip and 64GB total. AI5, AI6 and later generations are expected to feature a single chip per board, so all of that onboard memory feeds one SoC.

Memory supply has been tight across the industry due to ballooning AI demand, and “RAMageddon” isn’t showing any signs of letting up. DRAM maker Micron has estimated that each humanoid robot could need around 200GB of RAM. Against that backdrop, shaving gigabytes off each chip makes it easier to build more of them, but it also gives Tesla the challenge of having to make Optimus work with less than half the memory the industry estimates it will need.

Where AI5 Stands

Samsung recently began trial production of AI5 at its Taylor, Texas fab, with mass production aimed at 2027. On raw power, Musk has said a single AI5 chip delivers five times the compute of a full two-chip AI4 computer.

Don’t expect AI5 in your car anytime soon, however. Musk has said AI5 isn’t coming to vehicles first, with Optimus and Tesla’s AI training clusters getting priority before the hardware potentially trickles down into passenger cars.

Musk has long maintained that the current AI4 hardware is enough to support unsupervised autonomy. Tesla is already running driverless (and even steering wheel-less) Cybercabs and Model Y robotaxis without AI5, although the Cybercab does include more powerful FSD hardware than AI4 cars, even if that hardware isn’t AI5.

If Musk is right about bandwidth, the smaller memory pools shouldn’t hold Optimus back, and Tesla gets more chips out the door at a lower per-unit cost when volume production starts next year.