Nvidia unveils new AI GPUs: Blackwell Ultra, Vera Rubin & Rubin Ultra

Photo : Reuters

Nvidia has launched new AI-focused GPUs to keep its top spot in the tech world. These new chips aim to make computers faster and smarter. The company shared its plans at the GTC 2025 event, revealing its new GPU lineup for the next three years. The new models include the Blackwell Ultra GB300 (coming in 2025), followed by the Vera Rubin (2026) and Rubin Ultra (2027). Each of these GPUs will offer much more power and speed, helping businesses and tech firms handle heavy AI tasks faster.

Stay informed with the latest news. Follow DXB News Network on WhatsApp Channel  

Blackwell Ultra: More Power in 2025
The Blackwell Ultra GB300 is an improved version of Nvidia’s current Blackwell chip. It will be launched in the second half of 2025. While it keeps the same 20 petaflops of FP4 speed as the previous Blackwell model, it offers bigger memory and stronger performance.

Main features:

Memory: 288GB of HBM3e memory (more than the 192GB in the earlier Blackwell model).
Cluster Power: The DGX GB300 "Superpod" cluster will deliver 11.5 exaflops of FP4 speed and 300TB of memory (higher than the 240TB in Blackwell).
These upgrades will let companies process large-scale AI data much faster and more smoothly.

Vera Rubin and Rubin Ultra: Even Faster in 2026 and 2027
Nvidia will introduce the Vera Rubin GPU in 2026 and the Rubin Ultra in 2027. Both will bring huge speed gains and better performance for AI tasks.

Key features:

Vera Rubin:
50 petaflops of FP4 speed – 2.5 times faster than the Blackwell Ultra.
Designed to handle big AI models with ease.
Rubin Ultra:
Combines two Vera Rubin GPUs into one chip.
Offers 100 petaflops of FP4 speed and 1TB of memory.
It will be perfect for large-scale AI operations.
NVL576 Rack System:
Gives 15 exaflops of FP4 power and 5 exaflops of FP8 training speed.
14 times faster than Blackwell Ultra.
These new GPUs will cut down AI processing time massively. For example, Nvidia says the NVL72 cluster running the DeepSeek-R1 671B model will give AI responses in just 10 seconds—10 times faster than the H100 GPU from 2022.

New DGX Station: Strong AI Power for Small Teams
For smaller businesses or research teams, Nvidia has launched the DGX Station. This is a high-performance desktop designed for AI work.

Key specs:

Single GB300 Blackwell Ultra GPU
784GB of unified memory
800Gbps networking speed
20 petaflops of AI performance
This desktop will let smaller teams run heavy AI tasks quickly without needing big servers.

NVL72 Rack: Small Size, Big Power
Nvidia also revealed the NVL72 Rack, a compact but powerful AI unit. It offers:

1.1 exaflops of FP4 power
20TB of HBM memory
14.4TB/sec networking speed
This rack will help large tech companies process AI data at super-fast speeds, using less space.

Nvidia’s Growing AI Business
Nvidia’s latest move shows how much the company is focusing on AI technology. It has already shipped over $11 billion worth of Blackwell GPUs this year. So far, 1.8 million units have been bought by leading tech firms.

With AI demand rising, Nvidia’s data center business is now bigger than its gaming business, showing the shift in its focus. CEO Jensen Huang said the world now needs 100 times more AI power than last year, which shows how fast the tech is growing.

Future Plans: Feynman GPU in 2028
Looking ahead, Nvidia’s 2028 GPU will be called Feynman, named after the famous physicist Richard Feynman. This shows the company’s long-term commitment to AI innovation and its plans to stay ahead in the tech race.