Various market analyses$NVIDIA (NVDA.US)$GB200, Damo has brought in some new materials.

The GB200 supply chain has been launched, and it is believed that it will drive two new markets: chip testing and glass substrates.

Another thing worth noting is that Xiaomo continues to grow in depth in storage supply and demand, believing that the tense situation will continue until the second half of 2025, and the upward trend in prices will continue.

I. Damo: GB200 supply chain update

Recently, Damo updated the status of the GB200 supply chain, mentioning that the GB200 DGX/MGX supply chain has started and is currently in the design fine-tuning and testing phase. It is expected that orders and supplier allocations for 2025 will be finalized within the next few months.

1. Damo's forecast for GB200 shipments in 2024/2025

Based on CoWoS capacity allocation, Damo predicts:

• Approximately 420,000 GB200 superchips are expected to be delivered to the market in the second half of 2024;

• In 2025, GB200 chip production is expected to be about 1.5 million to 2 million units;

2. GB200 spawns two incremental markets: testing and packaging

Damo mentioned two incremental aspects brought about by GB200: semiconductor testing and semiconductor packaging.

• Semiconductor testing: In the GB200, Nvidia adopted a large chip strategy. The increase in chip size led to a decrease in yield, which in turn boosted demand for semiconductor testing. The data showed that Nvidia's AI GPU testing demand increased 20% month-on-month in 24Q2.

• Semiconductor packaging: The advanced packaging process adopted by GB200 will use glass substrates. This is mainly because compared with silicon and organic substrates, glass substrates have the advantages of adjustable strength, low energy consumption, and high temperature resistance. However, the downside is that the cost of using glass substrates is higher than that of silicon and organic substrates.

3. ASIC chips enter the AI semiconductor market

Damo predicts that ASICs (customized AI chips) will surpass the growth rate of GPUs in the next few years and are likely to occupy 30% of the cloud AI semiconductor market within four to five years.

The main reason is due to downstream demand for gradual personalization. Referring to HBM mentioned by Hynix and Micron at the beginning of this month, the gradual trend towards customization is also due to the personalized trend of downstream demand.

Similarly, ASICs are designed for specific AI tasks and can be optimized for specific algorithms and models to achieve higher performance than general-purpose GPUs.

Currently, major overseas cloud service providers, such as Google, AWS, Tesla, and Microsoft, are actively developing and deploying their own customized ASICs. The AWS Inferentia and Trainium chips mentioned in the report have 50% better performance per watt than GPUs.

2. Xiaomo: Storage supply and demand continue to be tight

In the second half of 2023, in order to meet the growing market demand for products such as HBM, DDR5, and QLC SSDs, storage vendors increased capital expenditure to expand production capacity.

The market believes that active supply from manufacturers will slow down memory price growth.

This time, Komo indicated in its report that despite the easing of the supply chain, the memory market will remain tight until the second half of FY2025. This is based on three factors:

1) The rise of AI terminals: AI phones and AI PCs will drive the growth of DRAM and RAM capacity in terminal devices.

The report predicts:

• The memory capacity of artificial smartphones will increase by 40%, and the average RAM will reach 28GB;

• Artificial intelligence PCs (mainly from HP, Lenovo, and Asus) will use 32GB of RAM;

• Smartphone/laptop content will grow at a CAGR of 14%/12% over the next three years;

2) Impact of yield loss: The yield loss in the CoWoS packaging process used for storage is high. This yield loss is about 10-15%, which will directly affect the effective supply of HBM.

3) Increased demand for NAND: Thanks to the adoption of QLC SSDs (Quad Cell Solid-State Drives) in AI servers. QLC SSDs will gradually replace traditional SSDs due to their high storage density and lower cost.

Based on these findings, the report increased the total memory target market (TAM) for the 2024/2025 fiscal year by 15%-18%.

Editor/Jeffrey