According to industry reports, currently most large models use NVIDIA's GPU chips. TrendForce predicts that global AI chip shipments will increase by 46% by 2023. NVIDIA GPUs are the mainstream choice in the AI server market, with a market share of about 60% to 70%.
Compared to the previous generation products, NVIDIA's new GH200 has basically the same "genes". It still features a 72-core Arm Neoverse V2 Grace CPU, Hopper GPU, and NVLink-C2C interconnect with a bandwidth of 900GB/s. However, the difference lies in the inclusion of the world's first HBM3e memory. The new chip no longer comes with the 96GB HBM3 vRAM and 480GB LPDDR5x DRAM found in the spring models of this year, but instead includes 500GB of LPDDR5X and 141GB of HBM3e memory, achieving a data throughput of 5TB/s.
In summary, this is the world's first chip equipped with HBM3e memory, which can increase local GPU memory by 50%. This is a special upgrade designed for the artificial intelligence market, as top generative AIs often have large sizes but limited memory capacity.
NVIDIA has stated that its latest GPU chip is equipped with the world's first HBM3e memory, which increases its local GPU memory by 50%. This is a specific upgrade for the artificial intelligence market, as top generative AIs often have large sizes but limited memory capacity.
According to the latest report from TrendForce, it is predicted that by 2024, market demand will significantly shift towards HBM3, while the demand for HBM2e is expected to decrease. Currently, SK Hynix is the main supplier of HBM3, while Samsung focuses on meeting orders from other cloud service providers. Micron plans to focus on developing HBM3e products, but its market share may slightly decline due to different customer compositions and the squeeze effect.
According to industry reports, Samsung plans to start production of HBM3 from the end of this year and plans to invest billions of Korean won to increase the HBM capacity at its Cheonan plant. Starting from the fourth quarter, Samsung's HBM3 will also be supplied to NVIDIA. Currently, SK Hynix exclusively supplies high-end GPU HBM chips to NVIDIA.
NVIDIA officials have stated that their latest GH200 product will not be in production until the second quarter of next year, and the price has not been announced yet. One important reason is that HBM3e will only be available for supply next year. Market reports indicate that Samsung and SK Hynix are expected to release HBM3E samples in the first quarter of next year and start mass production in the second half of 2024. Micron, on the other hand, chooses to skip HBM3 and directly develop HBM3e. By then, with the new NVIDIA chips, AI large models are expected to usher in a new round of breakthroughs.
Based on the observation of HBM supply and demand changes by TrendForce, it is estimated that by 2024, with various manufacturers actively expanding production strategies, the HBM supply-demand ratio is expected to improve, with an estimated change from -2.4% in 2023 to 0.6%.