Join Forward-thinking Leaders
Elevate your expertise with tech insights, startup breakthroughs, and leadership intelligence curated for your priorities.
AI Chip-Blackwell To Launch Soon by NVIDIA: Bold Move
NVIDIA is going to make a new AI chipset for China at a substantially cheaper price than its newly restricted H20 model, with mass manufacturing scheduled to begin in June.
Summary:
1. Nvidia is set to release a new AI chipset for China at a lower price than its H20 model, with mass manufacturing scheduled to begin in June.
2. The company’s market share in China has dropped from 95% before 2022, when US export barriers began, to 50% now, according to Nvidia CEO Jensen Huang.
3. The new GPU is expected to accomplish around 1.7 terabytes per second using GDDR7 memory technology, just under the export restriction limits.
Based on two sources, the GPU, or graphics processing unit, will be included in Nvidia’s current generation Blackwell-architecture AI processors and is estimated to cost between $6,500 and $8,000, which is far less than the $10,000-$12,000 the H20 sold for.
The new chip’s price, manufacturing timeline, and other features have not yet to have been disclosed. The three sources, media agencies talked with for this piece, declined to be identified because they were not permitted to speak to the media on this topic.
TSMC has declined to comment.
According to this AI tech news, China remains a major market for Nvidia, contributing to 13% of total sales in the previous fiscal year. It’s the third time Nvidia has had to design a GPU for the world’s second-largest economy due to restrictions imposed by US authorities that want to hinder Chinese technical growth.
Despite its substantially lower computing capability than the H20, Nvidia’s new AI Chip is expected to keep the corporation competitive in the face of significant market share losses owing to export restrictions. Its primary competitor in China is Huawei (HWT.UL), which manufactures the Ascend 910B chip.
Its “remaining edge lies primarily in its ability to integrate AI systems with its CUDA platform,” according to him.
The company’s programming architecture engineers utilise CUDA to construct AI chips, models, and applications on its GPUs. Its widespread use and the ecosystem that surrounds it encourage developers to continue with Nvidia.
Blackwell (AI Chip)
However, according to this latest AI tech news, Nicolas Gaudois, head of Asia technology research at UBS, believes that a new GPU with standard memory will be insufficient for some AI training and inference applications.
Nvidia’s market share in China has dropped from 95% before 2022, when US export barriers on its goods began, to 50% now, according to Nvidia CEO Jensen Huang, who spoke to reporters in Taipei last week.
Huang also advised that if US export restrictions continue, more Chinese consumers will purchase Huawei chips.
According to two sources, Nvidia is also creating another Blackwell-architecture processor for China, with production commencing as early as September. Reuters was unable to obtain the variant’s features quickly.
After the H20 was effectively banned in the United States in April, Nvidia considered manufacturing a modified version for China, according to sources, but that plan did not come to life.
Huang has said that the company’s earlier Hopper design, which powers the H20, can no longer accept additional improvements due to existing U.S. export limitations.
Reuters was unable to identify the official name for the new GPU, which is set to be released as early as June.
The most recent export limitations imposed new limits on GPU memory bandwidth, a critical criteria that determines data transmission speeds between the main CPU and memory chips. This capacity is especially relevant for AI tasks that require considerable data processing.
Jefferies predicts that the new restrictions limit memory bandwidth to 1.7-1.8 terabytes per second. That compared to the H20’s capability of 4 terabytes per second.
According to the latest AI tech news, GF Securities predicts that the new GPU (AI Chip) would accomplish around 1.7 terabytes per second using GDDR7 memory technology, which is just under the export restriction limits.
