Nvidia took the wraps off its A100 GPU, along with a raft of products that rely on the new GPU and new architecture. All of this should deliver massive improvements in both performance and power ...

Nvidia notes that this means nearly 500 times more memory than in a single Nvidia DGX A100 system. Let’s circle back to the 1 exaflop figure and break it down a little.

NVIDIA says every DGX Cloud instance is powered by eight of its H100 or A100 systems with 60GB of VRAM, bringing the total amount of memory to 640GB across the node.

That one didn't have a Nvidia Arm CPU and needed separate PCIe AI accelerators (A100); the 2025 iteration doesn't. It also carried a price of more than $100,000 at launch. Nvidia DGX Station ...

The new system is known as the Nvidia DGX GH200, ... It sounds like GPT-4 was trained on "tens of thousands of A100's" and in benchmarks H100 is about 4x faster at GPT-3 than A100 was.

The new AI supercomputer is equipped with 760 NVIDIA DGX A100 systems which it uses as compute nodes, offering a total of 6,080 NVIDIA A100 GPUs linked on an NVIDIA Quantum 200Gb/s InfiniBand ...

This provides 1 exaflop of performance and 144 terabytes of shared memory — about 500x more memory than the previous generation NVIDIA DGX A100, which was introduced in 2020, the company added.

Announced today at the company’s 2023 GPU Technology Conference, the service rents virtual versions of its DGX Server boxes, each containing eight Nvidia H100 or A100 GPUs and 640GB of memory.

The supercomp currently runs 760 Nvidia DGX A100 systems, comprising 6,080 GPUs, as its compute nodes. The GPUs communicate via an Nvidia Quantum 200Gb/s InfiniBand two-level Clos fabric.

NVIDIA says it created a supercomputer designed to help build generative AI models. The architecture of the DGX GH200 enables hundreds of powerful chips to act as a single GPU.

NVIDIA's new DGX Spark is a new mini supercomputer offering 1000 TOPS of AI power, with a price of $4000, and it launches at retail this month.