At the Huawei Connect Live 2019, Huawei announced the launch of its overall computing strategy for the first time. This means that computing strategy will become one of Huawei‘s core strategies in the future when the computing industry is turning to the turning point and computing architecture innovation is celebrating its golden decade. Huawei also released its Atlas 900 AI computing platform in the conference.
(Ken Hu outlines Huawei’s computing strategy in his keynote at HUAWEI CONNECT 2019)
So, do you know these Huawei Atlas AI services?
Huawei Atlas AI Computing Platform is powered by Huawei Ascend series AI processors and mainstream heterogeneous computing components. It provides a rich array of product form factors, such as modules, cards, edge stations, and appliances, enabling all-scenario AI infrastructure solutions for device-edge-cloud. It is widely applied in scenarios such as safe city, smart transportation, smart healthcare, and AI inference. As an important part of Huawei’s full-stack AI solution, the Atlas AI Computing Platform unlocks unparalleled compute power to help customers embrace an AI-driven future. (From Huawei)
Now Huawei has provided four products series for the market. If including the latest Atlas 900, there will be five series.
Atlas 200 AI Accelerator Module
Packaged in a form factor half the size of a credit card, the Atlas 200 consumes only 10 Watts of power, supports 16-channel real-time HD video analytics, and can be deployed on devices such as cameras and drones.
Atlas 300 AI Accelerator Card
With a HHHL PCIe standard card form factor, the Atlas 300 is positioned for data center and edge servers, supports multiple data precisions, and delivers up to 64 TOPS of INT8 on a single card, a 3x lead over competing products in the industry.
Atlas 500 AI Edge Station
This industry-leading edge product integrates AI processing capabilities and supports 16-channel HD video processing in the size of an STB, delivering a 4x performance over existing products in the marketplace.
Atlas 800 Deep Learning System
Based on the standard framework and programming environment, it provides an optimized AI environment with a pre-installed software library for cluster management and task scheduling, and system-level performance monitoring.
How about Atlas 900 AI Training Cluster?
Consisting of thousands of ascending 910 AI processors
The Atlas 900 AI training cluster, consisting of thousands of ascending 910 AI processors, is the fastest AI training cluster in the world. Its total power is 256P ~ 1024P FLOPS @FP16, which is equivalent to the computing power of 500,000 PCs.
At present, Huawei has deployed an Atlas 900 AI training cluster on Huawei Cloud, with a cluster size of 1024 Soaring 910 AI processors. Based on the current most typical ResNet-50 v1.5 model and the ImageNet-1k dataset, the Atlas 900AI training cluster can be completed in 59.8 seconds, ranking first in the world.
The “ImageNet-1k Dataset” contains 1.28 million images with an accuracy of 75.9%. Under the same accuracy, the test scores of the other two mainstream manufacturers are 70.2s and 76.8s respectively. The Atlas 900 AI training cluster is the same. 2 people are 15% faster.
Atlas 900 AI training cluster leading technology advantage
1.Atlas 900 is AI computing industry leader. The Atlas 900 AI training cluster uses the industry’s single-chip, the most powerful Ascendant 910 AI processor. Each 910 AI processor has 32 built-in DaVinci AI Cores. The single chip provides twice the computing power of the industry. 256TFLOPS@FP16). Followed by the best cluster network. The Atlas 900 AI training cluster adopts three types of high-speed interconnection methods: “HCCS, PCIe 4.0, 100G Ethernet”, and a hundred-terabyte full-connected non-blocking dedicated parameter synchronization network, which reduces network delay and shortens the gradient synchronization delay by 10 to 70%.
2.There are also system level tuning. The Atlas 900 AI training cluster integrates the high-speed interfaces of HCCS, PCIe 4.0 and 100G RoCE through Huawei’s collective communication library and job scheduling platform, fully releasing the powerful performance of the 910 AI processor.
3.The Atlas 900 AI training cluster uses a full liquid cooling solution to innovatively design the industry’s strongest cabinet-level closed insulation technology, supporting >95% liquid-cooled ratio. The single cabinet supports up to 50kW of ultra-high thermal power consumption, achieving the ultimate data center energy efficiency of PUE<1.1.
Atlas 900 AI cluster application scenario
The Atlas 900 AI cluster provides superior computing power for large-scale dataset neural network training. It can be widely used in scientific research and commercial innovation, enabling researchers to train AI models such as images, videos and voices more quickly. The mystery of the universe, forecasting the weather, exploring oil and accelerating the commercial process of autonomous driving.
For example, in the astronomical exploration, on a southern starry sky map with 200,000 stars, astronomers want to find a certain feature of the star, which requires 169 days of work. Atlas 900, it only takes 10 seconds.
The computing industry is a typical technology-intensive, capital-intensive industry. In the early stage, huge R&D investment is often required, and it is necessary to continuously maintain the ability to innovate and ensure competitiveness. At present, after the stability of the global computing industry pattern for many years, it has ushered in a new round of changes. Computational architecture innovation has entered a new golden age. The computing market will be a market with huge space and full of unknown competition.
For the first three decades, Huawei has focused on the issues between the connections and has achieved great success. Nowadays, in the era of intelligence, connectivity and computing are inseparable. Huawei chooses to start with the computing industry. Challenges and opportunities can coexist.
If you are interested in this Atlas 900, view these resources for more: