Nvidia Announces Blackwell Ultra NVL72 Deployments with Optical Interconnects
Cloud hyperscalers begin early access clusters featuring co-packaged optics, reducing inter-rack latency by 45% for multi-trillion parameter training runs.
- NVL72 racks transition from copper backplanes to co-packaged optical transceivers.
- Reduces inter-GPU latency across dual-tier clusters by 45%.
- Major cloud operators slated for volume delivery in Q1 2027.
Nvidia has formally initiated early delivery access for its flagship Blackwell Ultra NVL72 system configurations, signaling an aggressive architectural shift from copper backplanes toward co-packaged optical (CPO) interconnect fabrics.
The transition aims to eliminate the thermal and signal attenuation limits encountered when interconnecting thousands of B200 accelerators across hyperscale datacenter footprints.
Figure 1: Silicon die fabrication inspection and high-density copper-to-optical interconnect bridge.
Breaking the Copper Barrier in Rack-Scale Computing
While traditional copper direct-attach cables (DAC) have served as the backbone for dense compute clusters, multi-trillion parameter models require non-blocking all-to-all communications that push copper past its passive physical distance thresholds.
By integrating optical engines directly onto the package substrate alongside the NVLink 5 switches:
- Inter-rack bandwidth: Scales to an aggregate 1.8 TB/s bidirectional throughput per accelerator.
- Thermal dissipation: Interconnect-related thermal overhead dropped by an estimated 32% per rack footprint.
Cluster Switch Layer <---- 800G Optical Trunk ----> NVLink Switch Tray
│ (Co-Packaged CPO)
▼
Blackwell B200 DieProduction Timelines
Volume deliveries to Tier-1 cloud hyperscalers—including Microsoft Azure and Google Cloud—are scheduled to ramp in early 2027. Preliminary engineering samples indicate zero packet loss during sustained 72-hour synthetic all-reduce stress evaluations.