On December 26, the Financial Community reported (editor Shi Zhengcheng) that as 2025 approaches, the Global chip industry is eagerly awaiting the start of a new round of "arms races" — industry leader NVIDIA is set to rewrite the limits of chip computing power in the coming months.

Considering that a "back view image of a server" sparked the hype around the Copper concept when NVIDIA launched the Blackwell B200 chip earlier this year, the recent status of these computing power giants has been closely monitored by many investors.

For the NVIDIA stock price that has been fluctuating since the end of October, the importance of the launch of the new generation flagship products is also self-evident.

(Nvidia daily chart, source: TradingView)

The computing power is looming! 5090 PCB design leaked.

On January 7 next year at 10:30 AM Beijing time, Huang Renxun will announce the RTX 50 series graphics cards based on the BlackWell architecture during the opening speech at CES in Las Vegas. The first released graphics cards include the 5090 and 5080, and the 5070 Ti and 5070 may also be showcased, but their release will be delayed.

Although the parameters of the new graphics cards have been mentioned numerous times in various reports, a photo of the "5090 PCB board" on the technology forum Chiphell this week still provides a fresh perspective.

According to previous leaks, the chip area of the RTX 5090 graphics card, the GB202 chip, will reach 744 square millimeters, which is an increase of 22% compared to the 4090's AD102. The 16 pads around the GPU correspond to the mentioned 32GB video memory.

With the bandwidth improvement brought by the new generation of GDDR7 video memory, the industry is paying attention to the performance of this graphics card in high-load environments, such as 4K, 8K gaming, AI applications, and professional content creation.

In the trend observed over the past two years, NVIDIA is more inclined to enhance the performance limits in the 90 series. Leaks indicate that the RTX 5090 graphics card will have 21,760 CUDA cores, while the 5080 graphics card will feature 10,752 CUDA cores and 16GB of GDDR7 video memory.

The PCB design also shows that the new graphics card uses a 12V-2x6 power connector supporting up to 600W power, replacing the previous generation's 12VHPWR interface. Previously, after the launch of the 4090, users reported issues of burning or melting power interfaces.

The B300 ushers in a new round of upgrades for AI Chips.

While major AI companies are still waiting for the delivery of the B200 Server, NVIDIA's next-generation AI flagship chip is quietly approaching.

According to the latest leaks, NVIDIA will release the B300 chip and the corresponding GB300 Server platform at GTC 2025, which is set for late March next year. The B300 is the upgraded version previously known as "Blackwell Ultra."

Compared to the previous generation B200 chip, the design power consumption (TDP) of the B300 will increase from 1000W to 1400W. Accompanying this will be a comprehensive enhancement in architecture, configuration, and performance.

The most noticeable parameter upgrade of the GB300 is the memory. Compared to the GB200's use of an 8-layer stacked 192GB configuration, the new generation of AI Servers will utilize a 12-layer stacked 288GB HBM3e.

In comparison, the main competitor AMD recently released the MI325X, which offers 256GB of memory, and it is expected that the MI350X, set to be released in the second half of next year, will have similar parameters to the GB300.

Other upgrades of the GB300 include: the network card will use ConnectX 8, the optical module will upgrade from 800G to 1.6T, and the cooling system will also be redesigned to include more advanced water cooling plates. The cabinet will come standard with a capacitor tray and offer an optional battery backup unit system. Preliminary leaks indicate that through the upgrade of the Ultra architecture, there will be a 1.5 times performance improvement in FP4 for a single card. Additionally, the server computing board of the new platform will use LPCAMM memory modules for the first time.