The annual GPU Technology Conference (GTC) by NVIDIA is about to begin. After five years of waiting online, this technology feast is finally returning to the public eye in offline form, bringing unprecedented expectations to the industry.

The product release is undoubtedly the highlight of the GTC conference. According to industry sources, NVIDIA is expected to announce detailed information about its latest AI chip, the B100 Blackwell. This chip is expected to replace the existing H200 GPU and become NVIDIA's most powerful AI chip. Industry observers predict that the training speed of the B100 chip will be an order of magnitude faster than the current chip, thereby setting a new benchmark in the field of AI computing.

However, with the improvement in performance, power consumption has also attracted widespread attention. Although there are rumors that the B100 chip requires liquid cooling, Jeff Clarke, the Chief Operating Officer of NVIDIA's partner Dell, stated that although the chip has a power requirement of up to 1000 watts, which is 42% higher than the previous generation product, it does not require liquid cooling. This undoubtedly eliminates a major obstacle for the application and promotion of the chip.

It is worth mentioning that the B100 GPU will adopt advanced 2.5D packaging technology - TSMC's CoWoS-L, which stacks two computing chips together and connects 8 8-Hi HBM3e memory stacks, with a total capacity of 192GB. This design not only improves processing capabilities but also saves space and reduces power consumption.

In the future, NVIDIA seems to have even grander plans. It is reported that the next-generation Blackwell GPU, codenamed B200, is expected to achieve higher memory capacity of 288GB using 12-Hi technology, further consolidating NVIDIA's leading position in the field of AI computing.

For more specific details, stay tuned for tomorrow's report.