NVIDIA DGX-1 - innovate faster

The NVIDIA DGX-1 deep learning system comprises a combination of hardware and software that delivers faster and more accurate training of neural networks. Purpose built for deep learning and AI analytics, the DGX-1 delivers performance equivalent to 250 conventional CPU-only servers.

Configure DGX-1 Server

Performance

First generation DGX-1 with NVIDIA Pascal

The first generation DGX-1 is powered by eight NVIDIA Tesla P100 GPU accelerators, each with 3584 CUDA cores and 16GB of RAM. Unlike conventional GPU servers which use the PCI-E bus for communication between the host system and Tesla cards the DGX-1 uses NVLink, which is 5 to 12 times faster than PCI-E.

Second generation DGX-1 with NVIDIA Volta

The second generation of the DGX-1 is powered by eight NVIDIA Tesla V100 GPU accelerators which are based on the new Volta architecture. These cutting-edge GPUs combine both CUDA cores (5120) and the latest Tensor Cores (640) plus 16GB of RAM and are specifically designed for deep learning delivering a massive 5x speed up compared to the first-generation DGX-1. The second-generation DGX-1 is available to pre-order now with customers receiving a first-generation DGX-1 immediately, with a free upgrade to V100 GPUs later this year.

Save time and money

NVLink unlocks the full performance of the eight Tesla V100 cards, so the DGX-1 delivers up to 960 teraFLOPs at half-precision (FP16), the most common format used in deep learning calculations. This is a dramatic speed up compared to deep learning on CPUs, saving you significant time and money.

DGX-1 with V1007.4hrs
8x GPU Server18 hours
Dual CPU ServerX711 hours
Time to train ResNet50, 90 epochs to solution
DGX-1 with V100960 TFLOPS
DGX-1 with P100170 TFLOPS
Dual CPU ServerX3 TFLOPS
Half-precision (FP16) performance

Quad EDR IB

High-bandwidth and low-latency, with a total of 800GB/s of communication.

NVIDIA Tesla V100

The first GPU architecture to incorporate Tensor Core technology designed for AI.

Next Generation NVIDIA NVLink

High-speed interconnect 300GB/s per GPU, 10X faster than current PCI-E Gen3 x16 interconnections.

Two Intel Xeon CPUs

For boot, storage management, and deep learning framework coordination.

3U Rackmount Case

Fits in a compact rack space, using 3200 Watts of maximum system power.

Software

The DGX-1 is much more than a GPU server, it is a deep learning solution comprising a finely tuned combination of hardware and software. Running a GPU-optimised version of Ubuntu Server Linux, the software stack comprises drivers, the NVDocker container tool, deep learning SDK, NVIDIA Cloud Management Service plus NVIDIA DIGITS which is used to run deep learning frameworks such as Caffe, Torch, TensorFlow and many more.

The Operating System (Ubuntu Server Linux) is optimised to take advantage of the hardware and software features of the system and CUDA 8, especially with respect to memory management and hardware communications.

The deep learning frameworks provided with the system are especially optimised to take advantage of the NVlink communication links among other enhancements, in order to optimize multi-GPU communication in the system.

NVIDIA DGX-1 Clusters

If your workflow needs more compute performance multiple DGX-1 supercomputers can be linked together in a cluster. The DGX-1 hardware lends itself well to clustering as it is equipped with four 25Gb Infiniband EDR ports and two 10Gb Ethernet ports.

Clustering multiple DGX-1 servers requires high performance storage and networking to be able to deal with the high I/O demands of deep learning. Out of the box the DGX-1 runs the NFS file system however, it does support other file systems such as Lustre.

Building a DGX-1 cluster requires a head node/director to allocate jobs to the different nodes. We recommend using scheduling software such as Slurm on the head node/director.

Alternatively some deep learning software such as Caffe2 is designed to run in distributed mode, so you don’t need separate management software.

Whatever your requirements Scan’s data scientists and data centre engineers will be able to advise you on the best configuration for your cluster.

Specification

Generation First Generation DGX-1 with NVIDIA Pascal Second- generation DGX-1 with NVIDIA Volta
GPUs 8 x NVIDIA Tesla P100 8 x NVIDIA Tesla V100
CUDA Cores 28,672 40,960
Tensor Cores 0 5,120
GPU RAM 16GB next-gen HBM2 memory per GPU
CPUs 2 x Intel Xeon E5 2698 v4
CPU Cores 40 Phyisical, 40 HyperThreading
System RAM 512GB ECC Registered DDR4 2133MHz
Storage 1 x 480GB Intel S3610 SSD for OS, 4 x 1.92TB SSDs in RAID 0 for data
Network 2 x Intel 10 Gigabit LAN, 4 x Mellanox MCX455A-ECAT Infiniband EDR, 1xGigaBit management LAN
Operating System Ubuntu Linux Server
Power Supply Redundant 3200W
Form Factor 3U rackmount
Dimensions 444 x 866 x 131mm (WxDxH)
Weight 61Kg
Operating Ambient Temperature 10 - 30°C