AI/ML: Model Training
Unleash Full GPU Performance for Distributed AI Workloads.
LT3™ eliminates the transport bottlenecks in model training and inference, dramatically reducing idle GPU time, accelerating network throughput, and reducing overall training time. No hardware or fabric changes required.
Scale-Up. Scale-Out. Scale-Across.
How BitRipple LT3™ Removes Barriers to Blistering-Fast Model Training.
AI model training clusters depend on fast, synchronized communication between GPUs. Current architectures like RoCEv2 struggle with packet loss, flow collisions, and congestion, leading to idle GPUs and slower time-to-train.
No need for expensive and complex hardware solutions for scheduling to avoid contention, LT3™ is a software based solution using a real time erasure-coding resiliency protocol at the source. Combined with spraying packets granularly across all available paths, and decoding out-of-order without retransmissions, LT3™ achieves breakthrough results: lower collective completion time (CCT), higher GPU utilization, and faster model convergence.
Using this approach, we have achieved within 2% of the lower theoretical bound for collective completion time (CCT).
Measurable Impact.
BitRipple LT3™ slashes idle GPU time and improves throughput by tackling transport issues at the source.
-
38.3% lower CCT compared to RoCEv2 + ECMP
-
More compute, less wait
-
Performance within 2% of the theoretical minimum
Designed for Modern AI.
Multipath Packet Spray
Distributes packets across all paths evenly, avoiding ECMP hash collisions.
Loss-Tolerant by Design
Erasure-coded transmission eliminates retransmissions and maximizes throughput on all paths.
Drop-in Simplicity
Software-based and can be deployed at multiple layers in the infrastructure.
Transparent Scalability
Supports CPU/GPU offload, SmartNICs, and scales across geographically separated data centers.
Accelerate Your
AI Infrastructure.
LT3™ was built for scale – both Intra and Inter-datacenter. It's time to remove the final bottleneck in your AI model training stack.