The NVIDIA® DGX™ line is rightfully famous. It’s a supercomputer in a box with incredible power, especially for AI use cases. But we think you’ll find its cousin, the NVIDIA® HGX™ A100 GPU to be even more impressive because of the ROI it can bring you.
The HGX A100 is not the DGX’s little brother. There is nothing little about this platform. It's just different, and the things that make it different are the reasons why we feature the HGX for AI and HPC clients.
The NVIDIA A100 is a GPU based on the new NVIDIA Ampere GPU architecture. It provides up to 20 times higher performance over NVIDIA Volta. The A100 GPU has significantly more on-chip memory than Volta-based GPUs, to maximize compute performance, and it’s made to be impressively scalable.
What’s even more exciting is that it supports multi-instance GPU (MIG) capability, allowing you to securely partition your A100 GPUs into up to seven isolated GPU Instances for CUDA applications. This feature lets multiple users with separate GPU resources accelerate their applications without affecting other users, allowing organizations to dynamically adjust to shifting demands. And, you get all-to-all GPU communication at full NVLink speed, essentially making all your partitioned GPU instances behave as if they were a single node. Fantastic, right?
The A100 is available in 40GB and 80GB-high-speed HBM2 memory versions. The 40GB version has 1555 GB/sec of memory bandwidth—a 73% increase over the Tesla V100. And the A100 80GB has the world’s fastest memory bandwidth, at over 2 TB/s, so you can run the largest models and datasets.
And we’ll send it pre-loaded with a full AI software stack, including Ubuntu, PyTorch, TensorFlow, CuDNN, NVIDIA CUDA, and Keras, among others. That means that when we ship you a server based on the HGX A100, you have a turn-key, ready-to-run AI system.
Now you see why I say there is nothing little about this technology. In fact, if you check, you’ll see that the NVIDIA DGX has the same capabilities.
The NVIDIA DGX is a pre-configured solution where the processor and networking technologies have already been selected for you. The NVIDIA HGX is a building block, designed to give you the acceleration needed for AI.
In addition to the flexibility of not being locked into someone else’s hardware choices, you also save a great deal of money when you choose not to go with this famous supercomputer in a box.