NVIDIA's new chip "delayed"? Don't panic, the impact is not that significant

Author: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

NVIDIA's next-generation chip Blackwell seems to have encountered some minor troubles.

On August 1st, research firm ALETHEIA reported that NVIDIA's Blackwell chip may be delayed. Subsequently, on August 2nd, JP Morgan and Morgan Stanley each released their analysis reports. Overall, the impact of the delay of the Blackwell chip is not as bad as expected.

1. JP Morgan's View

• Reasons for Delay: JP Morgan believes there are three main reasons, with chip design issues being the core reason.

a. B100/B200 chip design issues;

b. CoWoS-L packaging yield issue: only 60%, much lower than the 90%+ level of CoWoS-S;

c. System-level issues: such as heating and liquid leakage;

• Supply Chain Impact: The delay of B100 is expected to last for one quarter, but the shipment time of GB200 remains basically unchanged.

2. Morgan Stanley's View

• Reasons for Delay: Morgan Stanley believes this is not a delay, but an improvement. NVIDIA aims to further enhance the stability of Blackwell through "redesign".

• Changes: Due to the CoWoS-L packaging yield issue, B200A will switch to CoWoS-S packaging.

• Supply Chain Impact: Although Blackwell paused production at TSMC for two weeks, with expanded capacity in the fourth quarter, the delay can be recovered and Blackwell can be delivered on time.

Overall, the delay of NVIDIA's chip is mainly due to chip design issues, low CoWoS-L packaging yield, and system-level issues (such as heating, liquid leakage).

3. Expert Analysis

Over the weekend, Ronghe Semiconductor CEO Wu Zihao, based on his years of Fab experience and supply chain information, wrote an article titled "The Full Record of the Blackwell Chip Fiasco of NVIDIA's Next-Generation Chip Blackwell", explaining in detail the reasons for this delay and the impact on the supply chain.

Wu Zihao stated that the "redesigned" chip this time is b102, which is the foundation of all Blackwell chips.

- b102: composed of one GPU die + 4 HBM3e;

- b100: composed of two b102s;

- GB200 motherboard: composed of two b100s + one Grace CPU;

- Servers: composed of motherboards and various system-level components (liquid cooling, copper cables, etc.);

The redesign of the underlying basic chip (b102 chip) may potentially affect the subsequent motherboards, servers, and even the entire supply schedule.

4. Expected Blackwell Delay

Previously, the market expected Blackwell to start production in the third quarter, mass shipments in the fourth quarter, and officially mass shipments of servers in the first quarter of 2025 Both JP Morgan and Eddie Wu believe that the chip issue may affect the planned production in the third quarter, but with TSMC's capacity expansion in the fourth quarter, the delay in the third quarter can be recovered. Overall, although Blackwell's shipments will decrease this year, it will not be by much, and it will not have a significant impact on server shipments in the first quarter of 2025.

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

Therefore, there are different interpretations in the market regarding TSMC's specific production schedule. If production starts in the second quarter and output gradually increases in the third quarter, this delay may indeed have some impact on the supply chain.

5. Other Issues

• CoWoS-L Packaging Yield Issue: JP Morgan stated that the CoWoS-L yield is only 60%, but after confirmation by Eddie Wu and the Nomura research team, the actual yield is over 90%. In addition, Fabless manufacturers like NVIDIA usually have a Plan B, as mentioned by Morgan Stanley, which will use CoWoS-S packaging (currently with a yield of 99%) to replace CoWoS-L without affecting server shipments.

• Heating Issue: Morgan Stanley believes that heating and high voltage issues are common during the introduction of new products and are not expected to have a substantial impact on mass production plans, with production progressing in an orderly manner.

• Leakage Issue: In the main components of liquid cooling such as the water-cooled plate, manifold, CDU, and quick connect device (QCD), leaks are most likely to occur at the quick connect device, but the specific situation will only be clearer after actual implementation.

6. Supply Chain Impact

• NVIDIA: Peak shipments of H-series servers are expected in the second half of 2024 for NVIDIA, with servers equipped with H-series chips being the main source of performance this year. There was originally no revenue expectation for Blackwell in the third quarter of this year, and Blackwell server shipments in the fourth quarter are also not high. Server shipments were originally planned for 2025, so the overall impact on NVIDIA's performance is not significant.

• TSMC: The expected reduction of 20,000 pieces in the originally planned production of CoWoS-L this year (reduced from over 40,000 pieces to 20,000 pieces) will be reflected in TSMC's performance in the fourth quarter or the first quarter of next year.

• Optical Modules: The B100 comes with an 800G optical module, but the shipment of servers with optical modules will not affect 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 severe as initially imagined. It will not result in months of supply chain delays, and the CoWoS-L packaging yield is not as low as rumored in the market. As NVIDIA mentioned in an interview, "Blackwell sample trials have already begun widely, and production is expected to increase in the second half of the year."