Invidia reveals a revolutionary of the future of artificial intelligence until 2028 …
In its annual GTC 2025 in California, NVIDIA, NVIDIA, announced a set of new chips designed to accelerate artificial intelligence, which includes the Ultra Rubin card which will be launched in the second half of 2027, and the Feynman card which should arrive in 2028.
The CEO of the company, Jin-Soun Hwang, revealed advanced specifications for these chips, explaining that it will contribute to the operation of intelligent robots and to the management of the operation of billions of digital agents.
The company has also underlined the next generation of its artificial intelligence processors, while “Invidia” plans to launch the Blackwell Ultra B300 chip in the second half of 2025.
This segment will characterize the presence of two graphic treatment units (GPU) on the same model, which gives it 15 “Petaflop” – and a unit of measurement of the computer at work and equal to a million billion operations per second – from heavy calculation performance in FP4 format for each segment.
Blackwell Ultra NVL72 to speed up artificial intelligence – NVIDIA
When you use it in the composition of NVL72, it will provide a computer capacity of up to 1.1 “eczaffal” – equal to a million teraflops – heavy reasoning performance in FP4 format, which is equivalent to 1.5 times the performance of blackwell B200 performance. Each B300 graphic processing unit will be equipped with 288 GB of HBM3E memory, against 192 GB in Blackwell B200.
Future cards
The company has also revealed a revolutionary generation of its next graphic processors, which bears the name of Vera Rubin, which was unveiled for the first time at the Computex 2024, and should be launched in the second half of 2026.
This name came after the famous astronomer Vera Rubin, and it will be equipped with a memory of dozens of turbocharbs, in addition to a dedicated central processing unit designed by invitedia called “Vera”.
According to “Inviteia”, Vera Rubin will obtain much higher performance compared to its predecessor, Grace Blackweell, in particular in the fields of formation of artificial intelligence models and the implementation of reasoning.
This segment contains two graphics processing units integrated into a model, which allows it to provide a performance of 50 bitaflobl to carry out inference.
When you use it in the composition of the architectural NVL144, the computer capacity will reach 3.6 eczelbl from the performance of the FP4 format, which is more than three times the performance of the Blackwell Ultra card, which provides 1,1 eczlogs in the composition itself.
Speed Speed will see and Vera Rubin Pine Pine
As for the CPU, it contains 88 dedicated arms core with 176 tracks and has a 1.8 TB / S facade, providing a high speed connection between the central processor and the Rubin graphic processors.
While in the second half of 2027, “Involes” will launch the ultra slice rubin, which is a more advanced version of Vera Rubin.
This segment will be based on the composition of NVL576, where each graphic processing unit contains four Ritkel size processors, which will allow it to provide 100 bitavlobl in formatting FP4 per processing unit.
In terms of IT in the formation of the complete NVL576 package, the ultra rubin will provide the power of 15 locks to carry out the inferiority of the FP4, and 5 eczlogs to carry out the formation of artificial intelligence models in FP8 format, which makes it four times stronger compared to the composition of the Rubin NVL144.
Each ultra rubin also contains 1 TB of HBM4E memory, while the complete package will include 365 high speed memory teraoctes.
Rubin Ultra NVL576 arrived in 2027 from Invidia – Nvidia
By 2028, “Inviteia” will unveil a mysterious generation of artificial intelligence cards under the name of Feynman, where the CEO of the company indicated that it will work with the Vera CPU processor instead of the expected Richard processor, indicating unexpected changes in the name of name, interior engineering and the speed of performance.
During his speech at the conference of developers of the GTC 2025, Huang established an ambitious roadmap for the future of artificial intelligence, explaining that development in this field will remain closely linked to the success of “Invidia”.
He explained that data centers have become artificial intelligence factories because they do not produce material objects, but rather produce data used to operate artificial intelligence models.
Huang expected the “Inviteia” slices to work quickly 10 billion digital agents, a number that reflects the rapid spread of artificial intelligence technologies in various industries.
And he added that 100% of “invited” engineers will be supported by artificial intelligence systems by the end of 2025, which reflects the deep integration between modern technology and the semiconductor development process.
