Powering Intelligence with
Heterogeneous Compute

GPU/CPU/DPU tri-architecture, elastic orchestration of millions of cores, unleashing extreme performance for AI and scientific computing

Unleashing the full potential of next-generation GPU architectures with extreme-bandwidth interconnects and linear scalability

• NVIDIA H200 full support with NVLink Switch ultra-high bandwidth interconnect
• Single-node HGX architecture, linear scaling to massive GPU arrays
• Mixed precision FP8/INT4, inference throughput boost by orders of magnitude

CPU, GPU, and DPU working in symphony, orchestrating workloads for maximum efficiency across diverse compute architectures

• CPU (Intel Xeon Platinum/AMD EPYC) + DPU (BlueField-3) collaboration
• SmartNIC offloads network stack, freeing substantial CPU compute
• PCIe 5.0 + CXL 2.0 memory pooling, breaking single-node capacity limits

Container-native GPU resource allocation with instant provisioning and intelligent multi-tenant workload isolation

• Kubernetes GPU Operator, instant container launch
• Multi-Instance GPU (MIG) slicing, multi-tenant concurrent isolation
• NVIDIA GPU Cloud registry, pre-optimized framework library

Domain-specific compute engines optimized for AI workloads, delivering breakthrough performance in specialized scenarios

• Tensor Core matrix ops, Transformer training acceleration by orders
• NVDecoder hardware decode, zero-CPU video stream processing
• Grace Hopper superchip, unified ultra-bandwidth CPU-GPU memory

Tri-architecture integration: GPU compute accelerators + CPU general processing + DPU network offload