On December 26th, Caixin News Agency reported that at the critical moment of 2025, the global chip industry is eagerly waiting for a new round of "arms race" to begin - industry leader NVIDIA will rewrite the upper limit of chip computing power in the coming months.
Considering that at the beginning of this year, when Nvidia released the Blackwell B200 chip, a "server back image" sparked speculation about the concept of copper cables, the recent situation of computing power giants has been closely monitored by many investors.
For Nvidia's stock price, which has been fluctuating since the end of October, the importance of launching a new generation flagship product is self-evident.
The computing power that is about to be unleashed! 5090 PCB design leakage
At 10:30 am Beijing time on January 7th next year, Huang Renxun will unveil the RTX 50 series graphics card based on the BlackWell architecture at the opening speech of CES in Las Vegas. The first release graphics cards, including 5090 and 50805070 Ti and 5070, may also make an appearance, but the release time will be later.
Although the parameters of the new graphics card have been repeatedly revealed by various sources, a photo of the "5090 PCB board" on Chiphell's technology forum this week still provides a fresh perspective.
According to previous reports, the GB202 chip used in the RTX 5090 graphics card will have an area of 744 square millimeters, an increase of 22% compared to the AD102 of the 4090. And the 16 pads around the GPU also correspond to the 32GB video memory mentioned in the leak.
Considering the bandwidth improvement brought by the new generation GDDR7 graphics memory, the industry is paying attention to the performance of this graphics card in high load environments, such as 4K, 8K gaming, artificial intelligence applications, professional content creation, etc.
From the trend in the past two years, Nvidia is more inclined to "stack materials" to challenge performance limits on the 90 series. The disclosure shows that the RTX 5090 graphics card will have 21760 CUDA cores, while the 5080 graphics card will be equipped with 10752 CUDA cores and 16GB GDDR7 memory
The PCB design also shows that the new graphics card uses a 12V-2x6 power connector that supports up to 600W power, replacing the previous generation's 12VHPWR interface. Previously, during the after-sales service of 4090, some users reported that the power interface was burnt or melted.
B300 rolls up AI chip for a new round of upgrade
While AI giants are still waiting for the shipment of B200 servers, Nvidia's next-generation AI flagship chip is quietly approaching.
According to the latest revelation, Nvidia will release the B300 chip and corresponding GB300 server platform at GTC 2025 in late March next year. B300 is an upgraded version previously known as "Blackwell Ultra".
Compared with the previous generation B200 chip, the design power consumption (TDP) of B300 will be increased from 1000W to 1400W. Along with it comes a comprehensive improvement in architecture, configuration, and performance.
The most obvious parameter upgrade of GB300 is video memory. Compared to GB200's 192GB configuration with 8-layer stacking, the new generation of AI servers will use 288GB HBM3e with 12 layer stacking.
As a comparison, AMD's recently released MI325X provides 256GB of video memory, and it is expected that the MI350X to be released in the second half of next year will provide parameters similar to GB300.
Other upgrades to GB300 include: the network card will use ConnectX 8, the optical module will be upgraded from 800G to 1.6T, and the cooling system will be redesigned with more advanced water-cooled boards. The cabinet will come standard with capacitor trays and provide an optional battery backup unit system. Preliminary revelations indicate that upgrading to Ultra architecture will result in a 1.5-fold increase in FP4 performance per card. In addition, the server computing board of the new platform will use LPCAMM memory modules for the first time.