NVIDIA releases the Ampere GPU Powered DGX A100 Supercomputing System
NVIDIA CEO Jensen Huang announced on 14 May 2020 the new Ampere GPU powered DGX A100. This monstrous GPU system will certainly rock the HPC market, with it’s whopping 5 PETAFLOPS (peak performance at FP16). It’s the follow-up model of the VOLTA powered DGX-1 (8 GPUs) and DGX-2 (16 GPUs) products.
The DGX A100 contains 8 next generation GA100 GPUs of 40GB memory each. The system contains many disruptive improvements:
- A total of 320GB of GPU memory and 5 PETAFLOPS of performance for increased workloads!
- The individual GPUs have 2x to 3x speed improvements in comparison with V100 depending on the specific workload, so the DGX A100 is across the board faster than DGX-2. In some specific cases the peak performance of the DGX A100 can be ten-fold, due to much more versatile hardware options (supported mixed precision settings)!
- HBM memory speeds of up to 1.5 TB a second
- 6.5 kWatt peak power usage (better performance per watt than DGX-2)
- 6U rack height
- A lot of internal IO bandwidth improvements, to leverage the increased compute capacity of the Ampere GPU.
- FP32 Tensor core capability.
- Tensor core support for double precision FP64.
- Ampere contains int8 capabilities: Ampere is a more general engine for all AI workloads (also inference)
- Memory bandwidth speeds of 1.5 TB per second
- Doubling of NVlink bandwidth in comparison with older generation DGX-1/2
- Better utilization of the GPUs: 1 GPU can be hardware partionned in up to 7 instances (real hardware IO partitioning, no time sharing of the GPU!). So you could configure the DGX A100 into 56 lightning fast physical GPUs with each 5.7 GB of Ram!
- Internal Nvme Storage upgradeable to 30 TB.
- Introduction of float16 as a format for deep learning workloads!
Conclusion: the DGX A100 is a game changer with essential new improvements, like GPU hardware sharing and int8 support. It is thus a more general powerhouse, which can be used in a lot of different deep learning challenges, not only for training, but also for large scale inference (for configured in a network with many instances). Watch NVIDIA’s CEO’s presentation on the DGX-A1OO or read about the specs in this pdf datasheet.
Robovision is NVIDIA’s only elite ATP partner in Benelux, certified to install DGX A100 in your (cooled) rack, place your order with us today, and the DGX A100 will be shipped to you next month. We can bundle your DGX A100 with a deep learning AI at scale orchestration system, connected to thousands of labelers and managed by a sound native Docker/Kubernetes environment for optimal scaling.