🔗Cisco and Nvidia Team Up for 12.8 Tbps USFP Module
Cisco and Nvidia's USFP could change data center cabling
TL;DR
Cisco and Nvidia are developing a 12.8 Tbps USFP module, packing 32, 400 Gbps lanes into a single pluggable. This could drastically reduce data center cabling complexity and power consumption.
Cisco and Nvidia are teaming up to create a 12.8 Tbps USFP module, packing 32, 400 Gbps lanes into a single pluggable. This could drastically reduce data center cabling complexity and power consumption. If you're dealing with high-density GPU clusters, this could mean fewer cables, less rack space, and lower power draw. The USFP promises up to 12.8 Tbps of Ethernet bandwidth, consolidating eight GPUs worth of cabling into a single fiber optic bundle. Each USFP pluggable will support up to 100 watts, and a single bad laser or DSP could take out far more than a single GPU.
Key Points
Cisco and Nvidia's USFP module packs 32, 400 Gbps lanes into a single pluggable, offering up to 12.8 Tbps of Ethernet bandwidth.
The USFP promises to consolidate eight GPUs worth of cabling into a single fiber optic bundle, reducing rack space and power draw.
Each USFP pluggable will support up to 100 watts, employing a two-phase insertion mechanism for electrical connection.
A single bad laser or DSP in a USFP module could take out far more than a single GPU, impacting data center reliability.
Cisco's Silicon One P200 routing chips are already being used by Microsoft to distribute workloads across multiple datacenters.
Why It Matters
If you're managing high-density GPU clusters, the USFP could reduce cabling complexity and power draw significantly. Each USFP pluggable supports up to 100 watts, and a single bad laser or DSP could take out far more than a single GPU. This impacts data center reliability and efficiency, making it crucial for teams dealing with high-density GPU workloads.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.