Americas

  • United States
by Sam Reynolds

Nvidia’s made-for-China chip delayed due to integration issues: Report

News
Nov 24, 20234 mins
CPUs and ProcessorsGenerative AITechnology Industry

Nvidia’s AI-focused H20 GPUs bypass US restrictions on China’s silicon access, including limits on-chip performance and density.

Chips Semiconductors manufacturing
Credit: Shutterstock

Nvidia has been forced to delay the launch of its H20 chip, an AI-focused GPU, which is compliant with the US restrictions on high-end chip exports, because of issues server manufacturers were having in integrating the chip.

Nvidia is in the process of launching three GPUs – H20, L20, and L2 – to be compliant with US export restrictions and maintain market access in China, avoiding limits on-chip performance and density set by the Biden administration.

This news comes after Taipei-based Digitimes first reported that the China-compliant GPUs might be delayed.

Nvidia’s H20 GPU is now expected in the first quarter of next year, possibly occurring in February or March, according to Reuters.

However, these modified chips do not necessarily rule out new rules from the US Bureau of Industry and Security, the exports rulemaker, which might see more silicon bans for Nvidia.

“There’s an ongoing risk that BIS could still implement sanctions on those [China export compliant] new products, the H20, L20, L2, in the next few months to quarters, which could influence Nvidia’s future revenues,” Ben Yeh, a Taipei-based analyst with Canalys said.

Packaging constraints with the H20 chip

While Reuters’ report highlighted the apparent challenges with the technical integration of the chip, Ian Cutress, chief analyst of More Than Moore, also noted that packaging constraints exist for the H20 because it uses the in-demand Chip on Wafer on Substrate (CoWoS)  technology, with expected capacity expansion by TSMC not until 2024. 

“With any hardware meeting the new standards, Nvidia has the capacity to ship anywhere from 10 to 10 million units, depending on manufacturing capabilities,” Cutress said. “Although packaging capacity is a limiting factor, it’s primarily an issue for the H20 model, which uses CoWoS and is currently in high demand.”

But as per Canalys’ Yeh, there are limits to this demand.

“While the AI server market has experienced extraordinary growth, we expect this upward trend to persist, though at a more tempered pace relative to the heights of 2023. The sector maintains strong momentum, yet it’s prudent to moderate expectations moving forward,” Yeh said.

This moderating tempo would ease demand for CoWoS, perhaps creating a glut, which according to Yeh, may eventually become oversupplied as multiple players are aggressively expanding capacity.

Even retail GPUs impacted by rules

Under the rules set out by the US Department of Commerce and Bureau of Industry Security, which measures the Total Processing Power and performance density data, Nvidia’s latest data center-grade GPUs and many of its high-end retail cards are prohibited from export.

Dell has reportedly restricted the sale of AMD’s latest Radeon RX 7900-series GPUs and Instinct AI and HPC GPUs to China and 22 other countries which the US has chilly relations with.

These restrictions have caused such a shortage of GPUs in the country that thousands of Nvidia GeForce RTX 4090 gaming GPUs are being modified and placed into AI server farms, Wccftech recently reported.  

Before the new rules were put in place, Nvidia reportedly had prioritized sending a significant number of GeForce RTX 4090 cards to China, contributing to global shortages and inflated prices.

The card, which has an MSRP of  $1599, is now retailing globally for close to $2000 new and $1850 used, according to tracking websites.