LAB

Atlas 800 Training Server

The Atlas 800 training server (model: 9000) is powered by the Kunpeng 920 and Ascend 910 processors. It features the industry’s highest computing density, ultra-high energy efficiency, and high network bandwidth. The server is widely used in deep learning model development and training scenarios and is an ideal option for computing-intensive industries, such as smart city, intelligent healthcare, astronomical exploration, and oil exploration.

Features

The ultimate computing density
- 2.56 PFLOPS FP16 in a 4U space
- 1.7x the computing density of industry peers
Superior perf./watt
- Supports air cooling and liquid cooling
- 2.56 PFLOPS/5.6 kW1 ultra-high energy efficiency,
- 1.3x that of its counterparts
High-speed network
- 8 100G RoCE v2 high-speed ports
- Slashes cross-server chip interconnect latency by 10-70%

Resources

https://github.com/cmcmicrosystems/Atlas-800

Specifications

Form Factor	4U AI server
Processor	4 Kunpeng 920 processors
Processor Memory	Up to 32 DDR4 DIMM slots, supporting RDIMMs Up to 2933 MT/s 32 GB or 64 GB per DIMM
AI Processor	8 Ascend 910 processors
HBM	8 * 32 GB
AI Computing Power	2.56 / 2.24 / 2 PFLOPS FP16

Applications

Deployed in data centers to enable AI training:

Model training
HPC
Smart city Smart healthcare
Astronomical Exploration
Oil exploration

Technical Support

Proposal issues in Ascend ModelZoo
Join Slack Channel

Who Can Access

Canadian Academic Subscribers

Contact Us

Dr. Yassine Hariri
Senior Staff Scientist
AI/ML and Embedded Systems

LAB Support

Does your research benefit from products and services provided by CMC Microsystems?

LAB