As Nvidia unveiled its next-generation artificial intelligence (AI) chip "Vera Rubin" at "CES 2026," the world's largest IT and consumer electronics show held in Las Vegas this month, the tech industry is focusing on the new technologies that will be integrated into it. With each new product launch, Nvidia has applied cutting-edge technologies ranging from memory chips to the communications methods that link chips. It is part of a strategy to lock in its lead by maximizing performance and power efficiency.
According to Nvidia on the 15th, Vera Rubin is the company's latest product and is scheduled for mass production this year and shipment to customers. Like the earlier-generation products Hopper and Blackwell, which were named after global scholars, the new chip is named after U.S. astronomer Vera Rubin, who found evidence of dark matter. Nvidia said the new product will outperform the previous-generation AI Semiconductor Blackwell by more than five times. In particular, Nvidia said it newly applied silicon photonics, an optical semiconductor technology, to maximize the performance of the new product.
◇ Boosting communication speed with light signals, cooling servers with water to cut power use↓
Servers equipped with Vera Rubin are set to adopt silicon photonics, the industry's first approach to transmitting data with light. This technology exchanges data with servers using light (optical signals) instead of traditional copper wiring, and is evaluated to deliver five times better power efficiency and more than 10 times faster transmission speeds. As AI Semiconductors process massive datasets and require faster data handling to raise AI performance, it is seen as optimized for that need.
Another differentiator is the adoption of a 100% liquid-cooling system to rein in power consumption. Unlike air cooling, which uses airflow to dissipate server heat, liquid cooling uses water to remove heat from servers. Water's thermal conductivity is more than 25 times higher than air, allowing rapid and efficient dissipation or absorption of heat generated by high-heat components in servers (CPUs, GPUs, etc.). It also consumes less power than air cooling, which runs large fans and circulates cold air.
Security performance has also been reinforced. Nvidia equipped Vera Rubin servers with a second-generation monitoring system. Server failures during AI model training can significantly affect learning. To prevent this, Nvidia applied a software system that, through the new monitoring framework, takes over the original task without shutting down the system when an issue occurs, preventing disruptions to AI computation.
◇ AI Semiconductor performance boost with HBM4, TSMC 3nm
The most notable change in this AI Semiconductor is the adoption of HBM4 (6th-generation HBM), a memory chip. High Bandwidth Memory (HBM) maximizes not only capacity but also the speed at which data is delivered to the processor that serves as the brain of the AI Semiconductor. Vera Rubin will be the first in the industry to use HBM4. To push AI Semiconductor performance to the limit, Nvidia is said to be continuously demanding ways to further enhance performance such as transmission speed, even after signing initial purchase agreements with Samsung Electronics, SK hynix, and Micron for first-batch supply.
The Vera Rubin platform boosted compute performance by applying TSMC's 3-nanometer (nm; 1 nm is one-billionth of a meter) process. Nvidia said graphics processing unit (GPU) performance improved fourfold compared with the prior generation. The number of GPUs required to train the same model was cut to one-fourth of the previous product, and the expense to realize AI compute performance fell to one-tenth of the existing expense, the company said.