news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

Nvidia's new generation chip Blackwell seems to have encountered some minor troubles.

On August 1, research firm ALETHEIA reported that Nvidia's Blackwell chip may be delayed. Then, on August 2, JPMorgan Chase and Morgan Stanley each published their own analysis reports. Overall, the impact of the Blackwell chip delay is not that bad.

1. Morgan Stanley’s view

• Reasons for delay:JPMorgan believes that there are three main reasons, among which chip design issues are the core reason.
a. B100/B200 chip design issues;
b. CoWoS-L packaging yield problem: only 60%, far lower than the 90% level of CoWoS-S;
c. System-level problems: such as heating and fluid leakage;
• Supply chain impact:The B100 delay is expected to last a quarter, but the GB200 shipment schedule remains largely unchanged.

2. Morgan Stanley’s view

• Reasons for delay:Morgan Stanley believes that this is not a delay, but an improvement. Nvidia hopes to further improve the stability of Blackwell through "redesign".
• Variety:Due to CoWoS-L packaging yield issues, B200A will switch to CoWoS-S packaging.
• Supply chain impact:Although Blackwell production was suspended at TSMC for two weeks, the delay can be recovered after the capacity is expanded in the fourth quarter, and Blackwell can be delivered on schedule.

In general, Nvidia's chip delays are mainly due toChip design issues, low CoWoS-L packaging yield, and system-level issues (such as heat generation, liquid leakage, etc.)。

3. Expert interpretation

Over the weekend, Ronghe Semiconductor CEO Wu Zihao combined his many years of Fab experience and supply chain information to write an article titled "A Complete Record of Nvidia's New Generation Chip Blackwell Failure", which detailed the reasons for the delay and the impact on the supply chain.

Wu Zihao said that the "redesigned" chip isb102, which is the basis of all Blackwell chips.

- b102: consists of one GPU die + 4 HBM3e;
- b100: composed of two b102;
- GB200 motherboard: consists of two b100 + one Grace CPU;
- Server: consists of a motherboard and various system-level accessories (liquid cooling, copper cables, etc.);

The redesign of the underlying base chip (b102 chip) willIt may affect subsequent motherboards, servers and even the entire supply progress.

4. Blackwell delays expectations

Previously, the market expected Blackwell to start production in the third quarter, mass shipments in the fourth quarter, and official mass shipments of servers in the first quarter of 2025.

Both JPMorgan and Wu Zihao believe that the chip problem may affect the original production schedule in the third quarter, but as TSMC's production capacity expands in the fourth quarter, the delay in the third quarter can be recovered. Overall, although Blackwell's shipments will decrease this year, they will not decrease too much, and will not have much impact on server shipments in the first quarter of 2025.

It is worth noting that, contrary to market expectations, Nvidia revealed at its first quarter earnings conference that Blackwell chips will begin production in the second quarter and gradually increase shipments in the third quarter, and it is expected to generate significant Blackwell revenue this year. Morgan Stanley's report also pointed out that this "redesign" is actually an improvement to the chip rather than a delay.

Therefore, the market currently has different interpretations of TSMC's specific production schedule. If production starts in the second quarter and gradually increases in the third quarter, then this delay may indeed have a certain impact on the supply chain.

5. Other issues

• CoWoS-L packaging yield issues:Morgan Stanley said that the yield rate of CoWoS-L is only 60%, but Wu Zihao and the Nomura research team confirmed that the actual yield rate is just over 90%. In addition, fabless manufacturers such as Nvidia usually have a Plan B. As Morgan Stanley mentioned, CoWoS-S packaging (currently with a yield rate of 99%) will be used to replace CoWoS-L, which will not affect server shipments.
• Fever problem:Morgan Stanley believes that problems such as heat dissipation and high voltage are common in the introduction of new products and are not expected to have a substantial impact on large-scale production plans. Mass production will proceed in an orderly manner.
• Liquid leakage problem:Among the main parts of water cooling, such as water cooling plates, branch pipes, CDUs and quick connectors (QCDs), leakage is most likely to occur at the quick connectors, and the specific situation will become clearer after actual introduction.

6. Supply chain impact

• Nvidia:The shipment peak of H-series servers will come in the second half of 2024, and servers equipped with H-series chips are the main source of Nvidia's performance this year. Blackwell had no revenue forecast for the third quarter of this year, and Blackwell server shipments were not high in the fourth quarter. Server shipments were originally scheduled for 2025, so overall they have little impact on Nvidia's performance.
• TSMC:The CoWoS-L originally planned for mass production this year is expected to be reduced by 20,000 pieces (from more than 40,000 pieces to 20,000 pieces), which will be reflected in TSMC's performance in the fourth quarter or the first quarter of next year.
• Optical module:B100 comes standard with 800G optical modules, but the optical modules are shipped with supporting servers, which will have no impact on the demand for 800G in 2024/2025.

Overall, the delay of Nvidia's Blackwell chip is mainly due to chip design issues, and the impact is not as serious as initially imagined. There will not be supply chain delays of several months, and the CoWoS-L packaging yield is not as low as market rumors. As Nvidia said in an interview: "Blackwell sample trials have begun widely, and production is expected to increase in the second half of the year."