We are at the dawn of a new age of intelligence, where AI and high performance computing (HPC) are transforming our world. From autonomous vehicles to global climate simulations, new challenges are emerging that demand enormous computing resources to solve. NVIDIA HGX-2 is designed for multi-precision computing—which combines the power of high-precision scientific computing using FP64 and FP32 with the speed of lower-precision AI computing with FP16 and Int8—to provide a single flexible and powerful platform to solve these massive challenges.
Accelerated by 16 NVIDIA® Tesla® V100 GPUs and NVIDIA NVSwitch™, HGX-2 has the unprecedented compute power, bandwidth, and memory topology to train these models faster and more efficiently. The 16 Tesla V100 GPUs work as a single unified 2-petaFLOP accelerator with half a terabyte (TB) of total GPU memory, allowing it to handle the most computationally intensive workloads and enable “the world’s largest GPU.”
AI models are exploding in complexity and require large memory, multiple GPUs, and an extremely fast connection between the GPUs to work. With NVSwitch connecting all GPUs and unified memory, HGX-2 provides the power to handle these new models for faster training of advanced AI. A single HGX-2 replaces 300 CPU-powered servers, saving significant cost, space, and energy in the data center.
HPC applications require strong server nodes with the computing power to perform a massive number of calculations per second. Increasing the compute density of each node dramatically reduces the number of servers required, resulting in huge savings in cost, power, and space consumed in the data center. For HPC simulations, high-dimension matrix multiplication requires a processor to fetch data from many neighbors to facilitate computation, making GPUs connected by NVSwitch ideal. A single HGX-2 server replaces 60 CPU-only servers.
NVSwitch enables every GPU to communicate with every other GPU at full bandwidth of 2.4TB/sec to solve the largest of AI and HPC problems. Every GPU has full access to 0.5TB of aggregate HBM2 memory to handle the most massive of datasets. By enabling a unified server node, NVSwitch dramatically accelerates complex AI and HPC applications.
The NVSwitch topology diagram shows the connection of two GPUs for simplicity. All 16 GPUs connect to NVSwitch in the same way.
| HGX-1 | HGX-2 | |
|---|---|---|
| Performance | 1 petaFLOP tensor operations 125 teraFLOPS single-precision 62 teraFLOPS double-precision |
2 petaFLOPS tensor operations 250 teraFLOPS single-precision 125 teraFLOPS double-precision |
| GPUs | 8x NVIDIA Tesla V100 | 16x NVIDIA Tesla V100 |
| GPU Memory | 256GB total | 512GB total |
| NVIDIA CUDA® Cores | 40,960 | 81,920 |
| NVIDIA Tensor Cores | 5,120 | 10,240 |
| Communication Channel | Hybrid cube mesh powered by NVLink 300GB/s bisection bandwidth | NVSwitch powered by NVLink 2.4TB/s bisection bandwidth |
NVIDIA partners with the world’s leading manufacturers to rapidly advance AI cloud computing. NVIDIA provides HGX-2 GPU baseboards, design guidelines, and early access to GPU computing technologies for partners to integrate into servers and deliver at scale to their data center ecosystems.
Learn about the latest in high performance computing, deep learning, and artificial intelligence.