Nvidia’s New AI Chip ‘Blackwell’ Sample Defect Rumor

I bought NVIDIA in ＂. Oh, my. ＂…[Hwang Jung-soo’s semiconductor issue] “Seohak ants” in a massive accident

Nvidia’s New AI Chip ‘Blackwell’ Sample Defect Rumor

Deformation Reports Bloomberg
Problem with customer usage samples
Foreign Ministry Says “Delays in Delivery for More Than 3 Months”

HBM3E SK Hynix Delivery for Blackwell
Samsung Electronics is also preparing for delivery
NVIDIA “Does as per year-end production plan”

According to foreign media reports, “Design defects” have been found in NVIDIA’s latest artificial intelligence accelerator Blackwell. AI accelerators are semiconductors that package graphics processing units (GPUs) and high bandwidth memory (HBMs) and are Nvidia’s main products. Blackwell AI accelerators, which will be released under the names of “B100” and “B200,” were scheduled to deliver tens of billions of dollars worth of AI to Microsoft, Google, and Meta, following the current flagship products of “H100” and “H200.”

Design flaws found in TSMC to make NVIDIA AI chips

“The delivery of Nvidia’s next-generation AI accelerators will be delayed for more than three months,” the U.S. information technology magazine “Deformation” reported on the 3rd (local time), citing two anonymous officials familiar with AI production. “It will affect customers who ordered tens of billions of won worth of MS, Google, and Meta.” Citing an MS official, he said, “NVIDIA informed Microsoft and other customers of the ‘design defect’ of advanced AI accelerators including Blackwell this week.”

Blackwell’s design problems were reportedly discovered by engineers at TSMC, a Taiwanese foundry (semiconductor consignment production) company in charge of production and packaging (a process that makes multiple chips operate on one chip). “There was a problem with the connection part of ‘GB200’ that was made by connecting Nvidia’s two Blackwell AI accelerators,” DeInformation wrote.

“It will not be easy to mass-produce Blackwell until the first quarter of 2025,” Deformation continued. Bloomberg, a global economic news agency, also cited the report by Deformation.

Earlier this week, Nvidia CEO Jensen Huang announced, “We have started providing samples of Blackwell AI accelerators to our customers.”

Blackwell is a next-generation AI accelerator unveiled by Nvidia in March. There are 208 billion transistors that are considered as indicators of semiconductor performance. It is more than twice the previous H-series AI accelerator. At the time, Nvidia CEO Jensen Huang explained, “With Blackwell, AI training performance can be improved four times, saving cost and energy.”

HBM3E for Blackwell SK Hynix to deliver

If Blackwell’s delay in delivery is true, there is growing interest in its impact on Korean memory semiconductor companies. Blackwell AI accelerator is scheduled to include 5th generation HBM (HBME3E) 8th layer and an upgraded version of Blackwell UltraN 12th layer. SK Hynix has been supplying HBM3E 8th layer products to Nvidia since the first half of this year. Samsung Electronics is undergoing quality tests from Nvidia to deliver HBM3E for Blackwell.

If Blackwell’s delivery is delayed, it is expected that it will have a negative impact on both SK Hynix, which switched its HBM production line to HBM3E, and Samsung Electronics, which is ready to mass-produce HBM3E. If Nvidia responds mainly to H100 and others, there is a possibility that demand for ‘HBM3’, the fourth-generation HBM that will be installed in the H series, will increase. Some analysts say that it will not be a big blow to SK Hynix because HBM3E will be included in the H200 that is currently released as an upgraded version of H100.

An official from the semiconductor industry analyzed, “It will not be a defect in HBM made by Korean companies because they mentioned the problem as a ‘design defect’ of the connection die in the deformation.”

Nvidia did not make an official statement on the report. According to Bloomberg, however, an Nvidia official said, “(Blackwell) production will be carried out as scheduled by the end of this year.” Officials from Google and Microsoft did not explain.

(