Nvidia today unveiled the A100 80GB GPU for the Nvidia HGXTM AI supercomputing platform with twice the memory of its predecessor. The new chip with HBM2e doubles the A100 40GB GPU’s high-bandwidth memory to 80GB and delivers more than 2TB/sec of memory bandwidth, according to Nvidia.
The A100 80GB GPU is available in Nvidia DGX A100 and DGX Station A100 systems, also announced today and expected to ship this quarter.
Systems providers Atos, Dell Technologies, Fujitsu, GIGABYTE, HPE, Inspur, Lenovo, Quanta and Supermicro are expected to begin offering systems built using HGX A100 integrated baseboards in four- or eight-GPU configurations featuring A100 80GB in the first half of 2021.
Building on the diverse capabilities of the A100 40GB, the 80GB version is intended for a range of applications with large data memory requirements. For AI training, recommender system models like DLRM have large tables representing billions of users and billions of products. A100 80GB delivers up to a 3x speed-up, designed for rapid retraining of models.
Nvidia said the A100 80GB also enables training of the largest models with more parameters fitting within a single HGX-powered server, such as GPT-2, a natural language processing model with “superhuman generative text capability.” This eliminates the need for data or model parallel architectures that can be time consuming to implement and slow to run across multiple nodes.
With its multi-instance GPU (MIG) technology, A100 can be partitioned into up to seven GPU instances, each with 10GB of memory, according to the company. “This provides secure hardware isolation and maximizes GPU utilization for a variety of smaller workloads,” Nvidia said in its announcement. “For AI inferencing of automatic speech recognition models like RNN-T, a single A100 80GB MIG instance can service much larger batch sizes, delivering 1.25x higher inference throughput in production.”
On a big data analytics benchmark for retail in the terabyte-size range, the A100 80GB boosts performance up to 2x, “making it an ideal platform for delivering rapid insights on the largest of datasets. Businesses can make key decisions in real time as data is updated dynamically,” Nvidia said.
For scientific applications, such as weather forecasting and quantum chemistry, the A100 80GB can deliver acceleration – Quantum Espresso, a materials simulation, achieved throughput gains of nearly 2x with a single node of A100 80GB.
“Speedy and ample memory bandwidth and capacity are vital to realizing high performance in supercomputing applications,” said Satoshi Matsuoka, director at RIKEN Center for Computational Science. “The NVIDIA A100 with 80GB of HBM2e GPU memory, providing the world’s fastest 2TB per second of bandwidth, will help deliver a big boost in application performance.”
Nvidia said the A100 80GB includes the following features of the Nvidia Ampere architecture:
- Third-Generation Tensor Cores: Provides up to 20x AI throughput of the previous Volta generation with a new format TF32, as well as 2.5x FP64 for HPC, 20x INT8 for AI inference and support for the BF16 data format.
- HBM2e GPU Memory: Doubles the memory capacity and is the first in the industry to offer more than 2TB per second of memory bandwidth.
- MIG technology: Doubles the memory per isolated instance, providing up to seven MIGs with 10GB each.
- Structural Sparsity: Delivers up to 2x speedup inferencing sparse models.
- Third-Generation NVLink and NVSwitch: Provides twice the GPU-to-GPU bandwidth of the previous generation interconnect technology, accelerating data transfers to the GPU for data-intensive workloads to 600 gigabytes per second.
The A100 80GB GPU is part of the Nvidia HGX AI supercomputing platform, which brings together Nvidia GPUs, NVLink, InfiniBand networking and an AI and HPC software stack.
