March 18th - NVIDIA's GTC 2024 kicks off tomorrow, and CEO Jensen Huang is gearing up for a keynote speech where the next-generation GPU architecture will be officially announced.
According to the latest leaks, both the B200 and RTX 50 series graphics cards will utilize this groundbreaking architecture, promising significant performance improvements brought by Blackwell.
Recent leaks suggest that the B100 will feature two chips based on TSMC's CoWoS-L packaging technology, connected to 8 stacks of 8-Hi HBM3e memory, totaling 192GB in capacity. But that's not all.
The leaks also mention that the next-generation Blackwell GP in the B200 will utilize 12-Hi stacks to achieve even higher capacity, with memory reaching up to 288GB, although it's uncertain whether it will utilize HBM4.
Earlier, Dell's COO Jeff Clarke leaked that NVIDIA is set to release B200 products featuring the "Blackwell" architecture by 2025, with power consumption potentially reaching 1000W.
In fact, the power consumption of the RTX 50 series based on this architecture doesn't paint a rosy picture. Previous reports suggest that the RTX 5090, equipped with a PCIe Gen 6 16-Pin power connector, could surpass 600W in power consumption, with high-end variants exceeding 450W and mid-range ones ranging between 300-350W.
So, with such power-hungry architecture, it wouldn't be surprising if the RTX 5090 outperforms the RTX 4090 by 60% to 70% (almost twice the performance)!
