A naukri.com initiative
Siliconangle
3w
131
Image Credit: Siliconangle
Nvidia debuts new silicon photonics switches for AI data centers
- Nvidia has unveiled a new collection of data center switches that are more efficient than current-generation hardware.
- The new switches were introduced at Nvidia's GTC event, where Jensen Huang discussed the Blackwell Ultra graphic card and Vera Rubin series AI chips.
- There are two product lines: Spectrum-X Photonics (Ethernet-based) and Quantum-X Photonics (InfiniBand-based).
- Both switch lineups utilize co-packaged optics (CPO) for increased power efficiency, being 3.5 times more efficient than previous products.
- Data center switches are responsible for transferring data between servers using fiber-optic networks for faster speeds.
- Nvidia's switches integrate transceivers into the chip, eliminating the need for standalone transceivers, reducing hardware costs.
- By placing processor and transceiver on the same chip, data transfer speeds increase with fewer lasers needed for data conversion.
- The new switches offer higher throughput, with the InfiniBand-based series shipping with 144 ports providing 800 gigabits per second.
- Nvidia's Ethernet-based series will provide up to 400 terabits per second throughput and is set to launch in 2026.
- The silicon photonics switches are based on TSMC's implementation of CPO, offering enhanced chip manufacturing and stacking capabilities.
Read Full Article
7 Likes
For uninterrupted reading, download the app