Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
© 2018 NVIDIA Corporation. All rights reserved. NVIDIA, the NVIDIA logo, Tesla, and NVLink are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.
RECORD PERFORMANCEThe HGX-2 platform is powered by NVIDIA NVSwitch™ which enables every GPU to communicate with every other GPU at full bandwidth of
2.4TB/sec to solve the largest of AI and HPC problems.
REDEFINING THE FUTURE OF COMPUTING HGX-2 multi-precision computing platform allows high-precision calculations
using FP64 and FP32 for scientific computing and simulations, while also enabling FP16 and Int8 for AI training and inference. This unprecedented versatility provides unique flexibility to support the future of computing.
0.5TBAggregate High-Bandwidth
GPU Memory
EXPLOSION OF NETWORK COMPLEXITYAI models are becoming increasingly complex and diverse, from translating
languages to autonomous driving. Solving these models requires massive compute capability, large memory, and extremely fast connections between the GPUs.
24XHigher GPU-to-GPU
Bandwidth*
* Compared to two HGX-1-based servers connected with 4x IB ports
2 PFLOPSTotal Compute
NVIDIA HGX-2FUSING HPC AND AI COMPUTING INTO A UNIFIED ARCHITECTURE
EMPOWERING THE DATA CENTER ECOSYSTEMNVIDIA works with a wide range of partners to deliver the ideal AI and HPC
solution. With HGX-2, they can now integrate a state-of-the-art platform into their servers to advance the data center ecosystem.
Convolutional Networks Sequence & Attention Networks Generative Adversarial Networksy1 y2 </s>
GPU8
GPU3
GPU2
GPU1
</s> y1 y3x3 x2 </s>
GPU2
GPU2
GPU1
GPU3
GPU8
Reinforcement Learning New Species
9X9 9X9
20
6
Wij = [8x16]
16
DigitalCaps
1010
||L2||
32
256 ReLU Conv1
PrimaryCaps
8
......
Encoder/Decoder
Concat
ReLu
Dropout
BatchNorm
Pooling
LSTM
WaveNet
GRU
CTC
Beam Search
Attention
3D-GAN
Attention Speech Enhancement
MedGAN ConditionalGAN
DQN Simulation
DDPG
Mixture of Experts Neural Collaborative Filtering
Block Sparse LSTM
SEE HOW HGX-2 CAN ACCELERATEYOUR AI AND HPC WORKLOADS.
www.nvidia.com/hgx
AI TrainingHGX-2 Replaces
300 CPU-Only Server NodesWorkload: ResNet50, 90 epochs to solution
CPU server: dual-socket Intel Xeon Gold 6140
HPCHGX-2 Replaces
60 CPU-Only Server NodesWorkload: MILC (particle physics HPC application)
CPU server: dual-socket Intel Xeon Gold 6140
100
200
300
0 Dual-SocketCPU
1
HGX-2
20
40
60
0 Dual-SocketCPU
HGX-2
60x300x
1
NVIDIA NVSwitchesDirect GPU-to-GPU Connection
Between All 16 GPUs
12
NVIDIA® Tesla® V100 GPUs0.5TB Memory
16