Atlas AI Computing Solution

Atlas AI Computing Solution

 

Model: 3000

Atlas 200 DK AI Developer Kit

» 16 TOPS INT8, 24 W
» 1 USB Type-C | 2 CCM ports | 1 x GE network port | 1 SD card slot
» 8 GB memory

Atlas 200 AI Accelerator Module

» 16 TOPS INT8, 8.5 W
» 16-channel real-time HD video analytics, JPEG encoding/decoding
» 8 GB memory | PCIe 3.0 x4 interface

Atlas 300 Inference Card

» 64 TOPS INT8/32 TFLOPS FP16, 67 W
» 64-channel real-time HD video analytics, JPEG decoding
» 32 GB memory | 204.8 Gbit/s memory bandwidth, supporting ECC
» PCIe 3.0 x16, HHHL card

Atlas 500 AI Edge Station

Intelligent edge
» Industry-leading edge product capable of AI processing
» Stable operation at -40°C to +70°C outdoors without fan modules
Superb capacity in a compact size
» 16-channel HD video processing in the size of an STB
» Performance 4x that of industry counterparts
Edge-cloud collaboration
» Collaboration with private and public clouds to push applications and update algorithms on
the cloud
» Unified device management and firmware upgrade on the cloud

Atlas 500 Pro AI Edge Server

Powerful computing
» Supports 4 Atlas 300 AI inference accelerator cards for 256-channel real-time HD video
analytics in a single server, supercharging inference in multiple scenarios
» Runs on the 64-core Kunpeng 920 processor to unlock superior computing power for applications
Ultra-high energy efficiency
» Leverages the multi-core, low-consumption architecture of Kunpeng processors to build an
energy-efficient AI computing platform in inference scenarios
» A single Atlas 300 card consumes only 67 W, accelerating AI servers with higher energy efficiency

Atlas 800 Inference Server

Powerful real-time inference
» Supports 2 Kunpeng 920 processors of multi-core computing architecture, accelerating
applications efficiently
» Supports 8 Atlas 300 AI accelerator cards for 512-channel real-time HD video analytics
Energy-efficient computing
» The multi-core and low-consumption Kunpeng platform builds an energy-efficient AI
computing platform
» A single Atlas 300 card consumes only 67 W power, accelerating AI servers with higher
energy efficiency

Atlas 800 Inference Server Model: 3010

Flexible configuration adaptive to various workloads
» Flexible configurations for SAS, SATA, NVMe, and M.2 SSDs
» LAN on motherboard (LOM) and FlexIO cards to provide various network ports
Intelligent video analytics

» 2 Intel® Xeon® SP Skylake or Cascade Lake processors
» 7 Atlas 300 AI accelerator cards for 448-channel real-time HD video analytics

Atlas 800 Training Server Model: 9000

Highest computing density
» 2 PFLOPS FP16 computing power in a 4U height, with computing density 1.5x that of the
industry
Ultra-high energy efficiency
» A single server supports air/liquid cooling, with an ultra-high energy efficiency ratio of 2
PFLOPS/5.5 kW
» The energy-efficient Atlas 800 server is ideal for deployment in enterprise equipment rooms
and high-density clusters
High-speed network bandwidth
» 8 x 100G RoCE v2 high-speed interfaces, reducing the inter-chip cross-server interconnect latency by 10–70%

Atlas 900 AI Cluster Model: 9000

Industry-leading computing power
» Consisting of thousands of Ascend 910 AI processors, the Atlas 900 AI cluster provides
256–1024 PFLOPS FP16 computing power, and delivers the highest performance by training
a ResNet-50 model in 59.8s.
Optimal cluster network
» Supports HCCS, PCIe 4.0, and 100GE RoCE high-speed interfaces, and vertically integrate
the communication library, topology, and low-latency network for linearity > 80%
Superior heat dissipation
» The 50 kW full liquid cooling system for a single cabinet supports a liquid cooling ratio of
greater than 95% and a PUE value less than 1.1, saving equipment room footprint by 79%.