NVIDIA's New "Nuclear Bomb": RTX 5090 Design Leaks 🚀

Time: Column:News views:164

As 2025 approaches, the global chip industry is eagerly awaiting the next "arms race" to kick off. Industry leader NVIDIA is set to once again redefine the limits of chip performance in the coming months.

Earlier this year, NVIDIA's release of the Blackwell B200 chip sparked a frenzy around copper cable concepts with just a single "server backplane image." Unsurprisingly, the latest developments from this computing powerhouse have been closely followed by investors.

For NVIDIA's stock price, which has seen fluctuations since late October, the release of the next-generation flagship product is undeniably critical.

NVIDIA daily chart, source: TradingView

(NVIDIA daily chart, source: TradingView)


Power at the Brink of Revelation: RTX 5090 PCB Design Leaks

At 10:30 AM (PST) on January 7th, 2025, NVIDIA CEO Jensen Huang will take the stage at CES in Las Vegas to unveil the RTX 50 series GPUs, based on the Blackwell architecture. The initial lineup will include the RTX 5090 and 5080, with potential appearances by the 5070 Ti and 5070, though their market release may follow later.

Although the specs for the new GPUs have been widely speculated, a leaked photo of the "5090 PCB board" on the tech forum Chiphell this week provides a fresh perspective.

NVIDIA's New

According to previous leaks, the RTX 5090 will feature the GB202 chip, with a die size of 744mm² — a 22% increase over the AD102 chip in the 4090. Surrounding the GPU are 16 solder points, corresponding to 32GB of VRAM, as earlier rumors suggested.

With the bandwidth boost from the next-gen GDDR7 memory, the industry is particularly interested in how the 5090 will perform under demanding workloads, such as 4K/8K gaming, AI applications, and professional content creation.

In recent years, NVIDIA has shown a clear tendency to push the performance envelope with its 90-series cards. The leaks suggest that the RTX 5090 will boast 21,760 CUDA cores, while the 5080 will feature 10,752 CUDA cores and 16GB of GDDR7 memory.

The PCB design also indicates a switch to a new 12V-2x6 power connector, supporting up to 600W, replacing the previous 12VHPWR interface. This comes after user reports of melted or burned connectors with the 4090 series.


B300 Ushers in a New Wave of AI Chip Upgrades 🌟

While AI giants are still awaiting shipments of B200 servers, NVIDIA's next-generation AI flagship chip is already on the horizon.

According to the latest leaks, NVIDIA will unveil the B300 chip and its associated GB300 server platform at GTC 2025 in late March. The B300, also referred to as "Blackwell Ultra," represents a significant upgrade over its predecessor.

The design power (TDP) of the B300 will increase from 1000W to 1400W, alongside comprehensive improvements in architecture, configuration, and performance.

One standout enhancement is in memory. While the GB200 platform used an 8-layer stack with 192GB of memory, the GB300 will adopt a 12-layer stack with 288GB of HBM3e memory.

For comparison, AMD's recently released MI325X offers 256GB of memory, and its upcoming MI350X, expected in late 2025, is projected to match the GB300’s capabilities.

Other GB300 upgrades include:

  • Network adapters: Upgraded to ConnectX 8

  • Optical modules: Boosted from 800G to 1.6T

  • Cooling systems: Redesigned with advanced water-cooling plates

  • Server cabinets: Equipped with capacitor trays and optional battery backup units

Preliminary leaks suggest that the Ultra architecture upgrade will deliver a 1.5x improvement in FP4 performance per card. Additionally, the server computation boards for the new platform will debut LPCAMM memory modules for the first time.

Stay tuned — NVIDIA is rewriting the rules of performance, and the competition is heating up! 🔥