Nvidia's GTC2024 opens tomorrow, and Huang Renxun has also prepared a keynote speech. The next generation GPU architecture will be officially announced by then. Judging from the latest details, B200 and RTX50 series graphics cards will use this architecture, and the performance improvement brought by Blackwell is also obvious.

According to the latest news, the B100 will use two chips based on TSMC's CoWoS-L packaging technology, connected to eight 8-HiHBM3e memory stacks, with a total capacity of 192GB, and this is not the only one.

The news also mentioned that the next generation BlackwellGP update of B200 will use 12-Hi to achieve higher capacity, with a video memory of 288GB, but it is not sure whether it is HBM4.

Previously, Dell Chief Operating Officer Jeff Clarke broke the news: Nvidia will launch the B200 product with the "Blackwell" architecture in 2025, and the power consumption may reach 1000W.

In fact, the power consumption of the RTX50 series based on this architecture is not too optimistic. It was previously reported that after the RTX5090 uses the PCIeGen6 16-Pin power supply interface, the power consumption can exceed 600W, while the high-end series can exceed 450W, and the mid-range power consumption is between 300-350W.

So under the power consumption of such an architecture, it is not surprising that the RTX5090 graphics card will be 60% to 70% (nearly double) faster than the RTX4090, right?