NVIDIA revealed its next flagship AI chip at GTC 2025. It is named Rubin Ultra to replace its current Blackwell Ultra. NVIDIA plans to launch this chip in late 2027. Each Rubin Ultra package connects four AI chips directly together, letting them share cache memory.
NVIDIA packed a massive 1 TB of HBM4e memory into the package. The company claims each chip will reach 100 petaFLOPs of FP4 performance. NVIDIA also introduced their new Vera CPU to replace the older Grace model. The Vera uses 88 custom Arm CPU cores NVIDIA created themselves rather than using standard designs.
Thanks to SMT technology, these special cores run 176 processes simultaneously. The chip transfers data extremely fast with its 1.8 TB/s NVLink C2C connection. NVIDIA already mapped out its future beyond Rubin. They decided to name their next architect, Feynman, after physicist Richard Feynman.
The company aims to release its first Feynman-based chips in 2028. This announcement gives us a clear view of NVIDIA's hardware plans for the coming years.
NVIDIA packed a massive 1 TB of HBM4e memory into the package. The company claims each chip will reach 100 petaFLOPs of FP4 performance. NVIDIA also introduced their new Vera CPU to replace the older Grace model. The Vera uses 88 custom Arm CPU cores NVIDIA created themselves rather than using standard designs.
Thanks to SMT technology, these special cores run 176 processes simultaneously. The chip transfers data extremely fast with its 1.8 TB/s NVLink C2C connection. NVIDIA already mapped out its future beyond Rubin. They decided to name their next architect, Feynman, after physicist Richard Feynman.
The company aims to release its first Feynman-based chips in 2028. This announcement gives us a clear view of NVIDIA's hardware plans for the coming years.
