THE FIRST UNIVERSAL UNIFIED SYSTEM FOR DATA ANALYTICS, TRAINING, AND INFERENCE
The NVIDIA DGXTM A100 is the ultimate instrument of AI researchers. Fully optimized and a state of the art system for AI with 5 petaFLOPS. The first one with a unified usage for data analytics, training, and inference. It is an elastic server architecture to scale-up and scale-out computing needs. The data center accelerator featuring eight A100 GPUs, eight NVSwitches delivering a cross sectional bandwidth of 4.8 TB/s, and six Mellanox ConnectX-6 for 200 GB/s network connectivity.
Key innovations offered:
TF32 - For the exponential growth of AI networks a new precision and increased training performance. Works like FP32 for AI with range of FP32 and precision of FP16.
MIG – Multi-Instance GPU capability to assign resources with a tailored size to each workload, up to 7 GPU partitions that run in parallel with guaranteed QoS.
20x speed up – 20x faster than Volta FP32, no code change required
Fastest I/O architecture - High Speed networking with eight single-port ConnectX-6 adapters for HDR IB or 200Gb/s Ethernet connectivity plus one dual-port ConnectX-6 Ethernet adapter for storage and networking.
DGX A100 is the homogenous infrastructure for heterogeneous workloads
Product Sheet See Store Tools
No media found...