The companies will package BitRipple’s LT3 GPU transport into deployment-ready solutions for AMD and NVIDIA GPU platforms

Sterling Computers Corp., a global solutions integrator serving government, education and commercial customers, and BitRipple, Inc., developer of the LT3 transport layer for GPU collective communication, today announced a strategic partnership to bring measurable efficiency gains to the infrastructure that trains and serves Super Intelligence (SI) and other large AI models.

In large AI clusters, GPUs exchange data through collective communication operations. When the network stalls, every GPU in the collective waits. Production measurements of training workloads presented publicly by industry practitioners put the network’s share of GPU wall-clock time at roughly a third on average, and higher on the most affected workloads. With GPU capacity reserved years in advance and data center capacity constrained in the near term, recovering that idle time is one of the fastest ways to add effective AI capacity without new construction.

LT3 addresses these issues all in software. LT3 runs over standard network fabrics, requires no new switch silicon, NIC firmware or fabric reconfiguration, and integrates beneath the NCCL collective library, so application code is unchanged. The same software also addresses Scale-Across links between data centers, where long round trips make every retransmission expensive. In results published in coordination with participants from Intel, Coreweave, and UC Berkeley, LT3 reduced Collective Completion Time (CCT) by 30 to 40 percent compared with RoCEv2 using ECMP routing, within about 2 percent of the theoretical lower bound. LT3 is built on the fountain codes invented by BitRipple co-founder and CTO Michael Luby.

Together, the companies will build LT3-integrated solutions for NVIDIA and AMD GPU platforms in Sterling’s Solution Center, leveraging the company’s elite technology partnerships, including Dell Technologies and AMD. Sterling will offer the solution as part of Sterling AI Factory and through its public sector contract vehicles to the operators, national laboratories, agencies and enterprises that build AI infrastructure.

“Every AI operator we talk to is capacity constrained, and a large share of the GPU time they already own is spent waiting on the network,” said Scott Brown, CEO of BitRipple. “LT3 recovers that time in software, on the network fabrics they already run. Sterling gives us the customer reach to turn measured results into solutions that customers can buy, deploy and have supported.”

“Our customers are under pressure to deliver more AI capacity from infrastructure they have already purchased,” said Chris Cyr, CTO of Sterling. “LT3 goes after one of the largest sources of idle GPU time via software, and it fits the NVIDIA and AMD platforms we already integrate without new NIC hardware or fabric changes. Working with BitRipple enables us to bring LT3 to our federal, public sector and enterprise customers as a tested, supported solution.”