news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhitong Finance APP learned that the two new most advanced AI chips of AI chip overlord Nvidia (NVDA.US) have "started off badly" due to design defects and encountered severe engineering obstacles during the chip manufacturing process. These AI chips based on the breakthrough Blackwell architecture are crucial for Nvidia to further expand its leading position in the artificial intelligence computing infrastructure market. According to people familiar with the matter, this delay has affected the original B100/B200/GB200 lineup of the much-anticipated Blackwell architecture chips announced by Nvidia in March. The B100 may be eliminated due to design defects and chip engineering problems and replaced by the slightly lower-performance B200A. Although the Blackwell delay caused Nvidia's stock price to plummet, Wall Street institutions such as Bank of America believe this is an excellent opportunity to buy at a low point.

Nvidia is redesigning a version of its AI chip, known as an artificial intelligence accelerator, to work better with data center infrastructure that is compatible with its H100 chip architecture based on its previous generation of AI chips, called Hopper, according to people familiar with the matter.

However, people familiar with the matter said that the data center hardware field targeted by B100 is only a relatively niche market segment. The real large-scale market lies in the market targeted by B200 AI GPU and GB200, which is based on two B200 chips and is equipped with NVIDIA's self-developed Grace CPU.

In addition, people familiar with the matter also said that due to some issues in chip manufacturing and CoWoS packaging technology support, the GB200 product that combines the Grace CPU with Nvidia's AI GPU will not be available in large quantities as quickly as the market had previously expected.

The news, which was first reported by the Information website, was revealed by people familiar with the matter. These chip manufacturing problems reflect the challenges Nvidia faces in accelerating its pace of innovation. CEO Huang Renxun is launching new chip design architectures and new chip connection technologies more quickly to maintain Nvidia's almost monopoly advantage in the core infrastructure of artificial intelligence. The company dominates the entire AI chip market, accounting for more than 90% of the market share, and AI chips are the core basic hardware behind generative artificial intelligence tools such as ChatGPT. With its almost monopoly advantage, Nvidia's performance and total market value have soared rapidly in the past two years. Its market value once surpassed Apple and Microsoft, becoming the "world's highest valued listed company."

NVIDIA has been deeply involved in the global high-performance computing field for many years, especially its CUDA computing platform, which is popular all over the world. It can be said to be the preferred software and hardware collaborative system in high-performance computing fields such as AI training/inference. This is also the core force that drives NVIDIA's AI chip market share to 90%. The CUDA computing platform is a parallel computing acceleration platform and programming auxiliary software exclusively developed by NVIDIA, allowing software developers and software engineers to use NVIDIA GPUs to accelerate parallel general computing (only supports NVIDIA GPUs, not compatible with mainstream GPUs such as AMD and Intel).

Nvidia shares plunge on Blackwell production delays

However, Nvidia declined to comment on recent media reports about the manufacturing problems of its Blackwell architecture chips. A company spokesperson said that it has begun to widely send Blackwell samples to customers, and that the market demand for its Hopper architecture AI chips remains very strong. Hopper architecture AI chips currently include H100 and H200, which has been mass-produced. Nvidia also said that the production pace of Blackwell architecture chips is expected to accelerate in the second half of the year.

Reports about Blackwell chip manufacturing challenges caused Nvidia's stock price to fall 6.4% on Monday after another correction last week, but the broader recession sentiment caused a plunge in technology stocks also affected the stock's performance. Meanwhile, Nvidia's strongest competitor in the AI ​​chip field, AMD (AMD.US), rose 1.8% against the trend when U.S. stocks plummeted. Investors hope that this may be an opportunity for AMD AI GPUs to gain more market share from Nvidia.

Nvidia's AI chips occupy a core position in the data center hardware of cloud computing giants such as Microsoft and Google. These cloud giants are Nvidia's largest customers. These cloud giants are currently spending tens of billions of dollars to build ultra-large-scale AI data centers around the world to meet the huge cloud computing resources needed by their global customers to deploy artificial intelligence technology.

In May, Nvidia CEO Huang Renxun said at an earnings conference that Blackwell architecture AI chips will soon be fully put into production and are expected to be delivered to cloud computing providers on a certain scale in the fourth quarter of this year. He predicted at the time that demand for the new series and its predecessor, the Hopper architecture, would continue to exceed supply. "We will see sales of the Blackwell architecture this year," Huang Renxun said in a earnings conference call with analysts. The company plans to release its next quarterly results on August 28, Eastern Time.

Nvidia's next-generation AI chip architecture - the Blackwell-based AI chip series will usher in a new ultra-high performance upgrade. Amazon, Dell, Google, Meta, Microsoft and other technology giants will configure a large number of Blackwell AI GPUs in their latest data center AI server systems. Wall Street analysts generally speculate that these technology giants' demand for Nvidia hardware will far exceed market expectations. Recently, industry insiders revealed that due to the extremely strong global demand for Nvidia's Blackwell architecture AI chips that are about to be mass-produced, Nvidia has significantly increased its AI chip foundry orders with chip foundry giant TSMC by at least 25%.

NVIDIA's most popular AI chip H100/H200 GPU accelerator is based on NVIDIA's breakthrough Hopper GPU architecture, providing more powerful computing power than the previous generation, especially in floating-point operations, tensor core performance and AI-specific acceleration. What's more important is that the performance of AI GPU based on Blackwell architecture is much higher than that of Hopper architecture. On the GPT-3 LLM benchmark with 175 billion parameters, the inference performance of GB200 based on Blackwell architecture is 7 times that of H100 system, and it provides 4 times the training speed of H100 system.

Bullishness on Nvidia remains the main theme on Wall Street, with strong calls for "buy on dips"

Although the delay may affect the chip manufacturing process and capacity planning of Nvidia's chip maker TSMC, forcing TSMC's tight CoWoS capacity to be transferred to chip giants such as AMD, most Wall Street analysts have taken a calm attitude and still insist on bullish Nvidia's stock price trend.

And based on the fact that NVIDIA will launch a new generation of Blackwell AI GPUs to the market before the end of the year and the demand for NVIDIA H100/H200 AI GPUs remains extremely strong, some Wall Street analysts expect that this will stimulate a new round of performance and stock price increases for NVIDIA. Therefore, they have become more bullish on NVIDIA's stock price trend in the next 12 months in recent days, believing that a new round of NVIDIA stock price increases is ready to go, and saying that this round of plunge is an excellent opportunity to "buy on the dip" of NVIDIA.

"Given the acceleration of innovation, capacity bumps are likely to continue," TD Cowen analyst Matt Ramsay said in a research note.

He stressed that even if there is a delay of several weeks, it may not have a significant negative impact on Nvidia's explosive revenue growth this year or long-term performance growth. However, Ramsey said that much depends on how Nvidia eventually fixes these problems and how quickly it delivers chips to major customers. The analyst reiterated Nvidia's "buy" rating and reiterated a 12-month price target of $165 (Nvidia's stock price closed down 6.36% to $100.45 on Monday).

Morgan Stanley recently stated that Nvidia hopes to further improve the stability of the chip and servers equipped with Nvidia AI chips by "redesigning" the Blackwell architecture AI chip. It is expected that due to the yield problem of CoWoS-L packaging, B200A will switch to CoWoS-S packaging. Morgan Stanley emphasized that although the Blackwell architecture chip may suspend production at TSMC for up to two weeks due to the redesign, TSMC's chip manufacturing and packaging capacity will be expanded in the fourth quarter. It is expected to recover the previously delayed capacity and finally deliver Blackwell to cloud giants on schedule. Morgan Stanley reiterated Nvidia's target price of up to $144 within 12 months.

Bank of America emphasized that Nvidia's recent stock price drop provides an excellent opportunity to buy on dips, and reiterated its target price of $150 and the chip industry's preferred investment target. Regarding reports that Nvidia's next-generation Blackwell chip will be delayed, Bank of America said that Blackwell may have to wait until the fourth quarter to be included in the performance, and even if the production capacity is delayed, it may not affect Nvidia's Blackwell-related revenue. Nvidia and TSMC may be able to work together to solve chip design and engineering problems in the short term. At the same time, Bank of America emphasized that the strong life cycle of Nvidia's previous generation Hopper products will continue to drive Nvidia's rapid performance growth, and Bank of America said that the challenge is not in market demand but on the supply side.