Nvidia’s New “Nuclear Bomb” Leak
Nvidia is set to release the RTX 50 series graphics cards based on the Blackwell architecture in two weeks. The PCB design of the 5090 graphics card surfaced this week. The global chip industry is eagerly awaiting a new round of “arms race,” and Nvidia will once again rewrite the upper limit of chip computing power. When Nvidia released the Blackwell B200 chip earlier this year, a “server back view” ignited the copper cable concept speculation, and its recent situation is closely watched by shareholders. For Nvidia’s stock price, which has been oscillating since the end of October, the launch of a new generation of flagship products is crucial.
The computing power on the verge of emerging! Leak of 5090 PCB design
At 10:30 am on January 7th, Jensen Huang will unveil the RTX 50 series graphics cards based on the Blackwell architecture in his opening speech at CES in Las Vegas. The initial release will include the 5090 and 5080. The 5070 Ti and 5070 may also make an appearance but will be released later. Although the parameters of the new graphics card have been frequently leaked, a “5090 PCB board” photo on the technology forum Chiphell this week still provides a new perspective.
The GB202 chip used in the RTX 5090 graphics card has an area of 744 square millimeters, an increase of 22% compared to the AD102 of the 4090. The 16 solder pads around the GPU correspond to 32GB of video memory. Considering the bandwidth improvement of the new generation GDDR7 video memory, the industry is paying attention to the high-load performance of this graphics card, such as 4K and 8K games, artificial intelligence applications, and professional content creation. In the past two years, Nvidia has tended to “pile up materials” in the 90 series to challenge the performance limit. Leaks show that the RTX 5090 graphics card will have 21,760 CUDA cores, and the 5080 graphics card is configured with 10,752 CUDA cores and 16GB of GDDR7 video memory.
The PCB design shows that the new graphics card uses a 12V – 2×6 power connector that supports a maximum power of 600W, replacing the previous generation 12VHPWR interface. After the 4090 was released, some users reported that the power interface was burned or melted.
B300 Triggers a New Round of Upgrades in AI Chips
While AI giants are waiting for the shipment of B200 servers, Nvidia’s next-generation AI flagship chip is approaching. At GTC 2025 in late March next year, Nvidia will release the B300 chip and the corresponding GB300 server platform. The B300 is an upgraded version of the previous “Blackwell Ultra.”
Compared with the previous generation B200 chip, the design power consumption of the B300 is increased from 1000W to 1400W, bringing all-round improvements in architecture, configuration, and performance. The most obvious parameter upgrade of the GB300 is the video memory. Compared to the 8-layer stacked 192GB configuration of the GB200, the new generation of AI servers will use 12-layer stacked 288GB of HBM3e. For comparison, AMD’s recently released MI325X provides 256GB of video memory, and the MI350X expected to be released in the second half of next year can provide similar parameters to the GB300.
Other upgrades of the GB300 include: the network card will use ConnectX 8, the optical module will be upgraded from 800G to 1.6T, the cooling system will also be redesigned, and a more advanced water cooling plate will be added. The cabinet will be equipped with a capacitor tray as standard and an optional battery backup unit system will be provided. Preliminary leaks show that through the Ultra architecture upgrade, there will be a 1.5 times improvement in single-card FP4 performance. The server computing board of the new platform will use LPCAMM memory modules for the first time.
