NVIDIA Battles Manufacturing Woes with Blackwell GPUs, Delays Full-Scale Production to Q4 2024

As of last week, NVIDIA (NVDA) shares climbed 1.51%, though they dropped by 7.73% over the past week. Year-to-date, the stock has surged by 141.11%, positioned at a market capitalization of approximately $2.93 trillion.

Recent performance reviews surfaced concerns about the manufacturing challenges of NVIDIA’s advanced chips, particularly focusing on the Blackwell architecture GPUs. Despite reporting robust quarterly sales and profits, the company highlighted that manufacturing difficulties have led to reduced profit margins. Consequently, a provision of $908 million was recorded for the last quarter, causing the stock to drop 6.4% on Thursday.

NVIDIA acknowledged yield issues with the Blackwell architecture GPUs, necessitating redesigns of parts of the B200 processor to enhance yield rates. Consequently, full-scale production of the next-generation Blackwell GPUs will be delayed until the fourth quarter of 2024. The adjusted design for better production yields is planned to result in multi-billion-dollar revenues starting from the fourth quarter.

The engineering complexities are reportedly due to the immense size and intricate design of the Blackwell chips. These chips utilize TSMC’s advanced packaging technology, CoWoS-L. NVIDIA's intervention aimed at increasing yield, involving modifications to the top metal layer and bumps on GPU chips, is a response to significant packaging and production challenges.

To maintain its dominance in the AI chip sector, NVIDIA has employed a 'bigger-is-better' strategy, resulting in sizable yet difficult-to-manufacture chips. The newest AI chip, Blackwell, described as "massive," comprises two interconnected dice using TSMC’s 4nm process, featuring 208 billion transistors, which is 2.6 times that of its predecessor.

Despite adjustments, industry analysts predict persistent challenges due to the complex nature of TSMC’s advanced packaging technology and the inherent difficulties arising from the chip's size. G. Dan Hutcheson from TechInsights remarked that synchronizing the operation and yield rates of different chip parts is crucial, as low-yielding components could rapidly degrade the overall quality.

Striving to enhance yield rates, NVIDIA has adopted modifications and plans sequential production ramp-up beginning by the fourth quarter. These efforts are indicative of the broader industry challenge where chip manufacturers, aiming to enhance processing power, face increased production complexity.

Amid these challenges, competition from firms like AMD and Cerebras Systems has intensified. AMD’s CEO Lisa Su noted that escalating chip sizes lead to inevitable manufacturing complexities. Conversely, Cerebras targets the market with its sizable single-die chips aimed at AI cloud computing, cautioning that integrating multi-chip solutions presents exponentially growing difficulties.

Despite the current hurdles, NVIDIA is slated for ongoing growth driven by un satiable GPU demand, recently seen with improved orders for H200 and H100 chips. As NVIDIA continues to navigate manufacturing obstacles and sector-specific challenges, its future trajectory, particularly in revolutionizing AI chip technology, will be closely observed by industry stakeholders and investors alike.