NVIDIA H200 Ships in Q3, B100 Expected Next Year

It's reported that NVIDIA's AI GPU H200 has entered mass production in late Q2, with significant shipments expected in Q3. However, the accelerated launch of NVIDIA's Blackwell platform by one to two quarters may influence customer interest in the H200.

Current orders are predominantly focused on the H100 within the HGX architecture, with limited demand for the H200. The H200 units produced in Q3 will primarily support NVIDIA's DGX H200 systems. Some visibility exists for the B100, with shipments anticipated in the first half of next year.

The H200, an upgrade to the H100 GPU, is based on the advanced Hopper architecture and introduces HBM3e high-bandwidth memory technology, offering faster data transfer speeds and larger memory capacity. This makes it particularly beneficial for large language model applications. According to NVIDIA, the H200 improves generative AI response times by up to 45% compared to the H100, especially for complex models like Meta's Llama2.

Positioned as a milestone in NVIDIA's AI computing portfolio, the H200 not only builds on the strengths of the H100 but also achieves significant advancements in memory performance. With the commercialization of the H200, the demand for high-bandwidth memory is expected to increase, driving growth in the AI hardware supply chain, particularly for HBM3e suppliers.

The B100 GPU will feature liquid cooling technology, marking a significant shift in chip performance enhancement. With the H200 GPU's TDP at 700W and conservative estimates placing the B100's TDP near a kilowatt, traditional air cooling may no longer suffice, necessitating a transition to liquid cooling.

NVIDIA CEO Jensen Huang stated that starting with the B100 GPU, all future products will transition from air cooling to liquid cooling. According to Galaxy Securities, the B100 GPU is expected to deliver at least twice the performance of the H200 and over four times that of the H100. This performance boost stems from both advanced manufacturing processes and the crucial role of cooling technology. As chip performance demands increase, liquid cooling will become essential to meet the thermal management needs of high-power GPUs.