Container-level Fractional GPU Scaling
Assigning slices of SMP / RAM to containers
Shared GPUs: inference & education workloads
Multiple GPUs: model training workloads
With a proprietary CUDA virtualization layer
* Registered patent in Korea, US and Japan
Proprietary CUDA virtualization layer
- Supports all GPU models for CUDA 8 to 12(desktop / workstation / datacenter)
- No code change required for user programs
- No customization/rebuild required for DL frameworks
- It is not limited to TensorFlow/PyTorch; any GPU-accelerated computing workload works!
- Supports multi-GPU for single container using multiple fractions from different GPUs
- Reproducible R&D environments for faster experiment cycles
- On-demand resource provisioning on top of bare-metal, VMs, and containers
- Optimized for clusters of high-end nodes with many CPUs and accelerators