Nvidia has released a minimize-down version of its excessive-finish A100 GPU as a way to get around restrictions the US government lately imposed to limit how companies can sell AI training equipment to international locations treasure China and Russia. The card, acknowledged as the A800, may detached have the same massive computing capabilities as the original nonetheless can finest communicate with various cards in a computing cluster at 400 gigabytes per 2d instead of 600GBps. According to The Wall Road Journal, that means the A800 wants to be correct as apt for easy tasks nonetheless will probably be significantly less effective when it’s part of something treasure a supercomputer.
An unnamed Nvidia spokesperson told Reuters that the card is an “alternative product to the Nvidia A100 GPU for customers in China” that “meets the U.S. Authorities’s clear take a look at for diminished export defend an eye on and cannot be programmed to exceed it.” While the exact specifications the US exams against aren’t public, the WSJ does say they involve interconnect speeds, as neatly as “various performance thresholds.”
Nvidia’s demonstrate about no longer being able to be programmed to exceed the limits is important. Back when GPUs have been in fast supply, the company announced that its original cards would have limiters that couldn’t be hacked, supposedly making the cards less desirable for crypto mining. Famously, it wasn’t prolonged ahead of individuals have been able to get around those limiters.
The US has been putting heavy limits on what kinds of chips American companies can sell to China. In October, the Biden administration announced original restrictions on what kind of chips and chip manufacturing equipment may be provided to the nation, requiring manufacturers to obtain an export license. In September, Nvidia and AMD both warned investors that their businesses may be impacted by original restrictions on exporting their top-tier merchandise.
The federal government has said the limits are an attempt to limit China’s technical and military capabilities. Some technologists, treasure ragged Google CEO Eric Schmidt, have warned that the US isn’t capable of keeping up, although various researchers have pushed back against the arms race framing. It’s miles imaginable, on the opposite hand, that chips treasure the A100 may be venerable for military or various nefarious capabilities. As my colleague James Vincent pointed out, they’re going to be venerable to create “automated targeting techniques for drones or the facial recognition techniques China uses to surveil its population.”
Whether or no longer or no longer the A800’s slower interlink speeds will probably be ample to hinder or meaningfully slack down that work, it does seem treasure Nvidia thinks it’s ample to meet the federal government’s requirements. For the company, at least, that’s apt information — it estimated that the restrictions may affect up to $400 million in quarterly sales “if customers carry out no longer want to purchase the Company’s alternative product offerings or if the [US Government] does no longer grant licenses in a neatly timed manner.”
Correction November ninth, 1: 48PM ET: A outdated version of this article venerable the incorrect abbreviation for gigabytes per 2d, making it seem as if the A100’s interlink ran at a powerful-slower 600 gigabits per 2d. We remorse the error.
