""Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof," said Nvidia CEO Jensen Huang."
"The platform combines six chips into a single integrated system: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch."
"According to Nvidia, Rubin delivers up to 10 times lower cost per token than the Blackwell platform when inferencing mixture-of-experts (MoE) models."
"In addition, MoE models can be trained with 4 times fewer GPUs compared to its predecessor."
Nvidia's Rubin platform is a new-generation AI infrastructure composed of six integrated chips designed to lower costs and accelerate AI adoption. First Rubin systems will be available in the second half of 2026 through AWS, Microsoft, and Google. The platform integrates the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch. Rubin claims up to 10× lower cost per token for MoE inference and enables MoE training with four times fewer GPUs. Key innovations include NVLink 6 bandwidth, an 88-core Vera CPU, a third-generation Transformer Engine, rack-scale confidential computing, and a second-generation RAS Engine.
Read at Techzine Global
Unable to calculate read time
Collection
[
|
...
]