CLOUD DATA CENTER SOLUTIONS

Pain points

Inefficient Training Caused by Packet Loss

• In the AI computing environment, the RDMA network is often introduced to improve GPU training and inference efficiency. The packet loss rate greater than 0.1% causes a sharp decrease in the effective network throughput. The packet loss rate of 2% reduces the RDMA throughput to 0. To prevent the RDMA throughput from being affected, the packet loss rate must be lower than 0.001%, and preferably 0%.

Complicated RDMA Network Configuration

• RDMA network configuration involves dozens of parameters related to thresholds and optimization. It takes at least one month to adjust the network configuration before a cluster can be deployed. Can there be an easier way to configure the thresholds to quickly deploy an RDMA network?

Solution

Uses PFC and ECN based on the DC switch and RoCEv2 protocol to dynamically carry out traffic backpressure, achieving zero packet loss over the network and improving the efficiency of AI training.

Benefits

Efficient Training

• The backpressure technology is adopted to meet the requirement for no packet loss in the RDMA network and improve the training efficiency of the AI scenario.

Use Without Learning

• With the built-in scripts, users simply need to make a few clicks without learning to deploy the RDMA network, which is easy and fast.

Pain points

Under the background of the "carbon peaking and carbon neutrality" policy, it has become a general trend to build intensive, green, and super-scale DCs. Traditional air cooling cannot meet the PUE requirements of future DCs and the heat dissipation requirements of a single chassis.
The performance of switching chips, SerDes, and optical modules is constantly improving with the iteration of network architecture, and power consumption of DCI becomes a great challenge.
In a large DC, IT equipment accounts for over 50% of the energy consumption. Reduction in energy consumption, efficient heat dissipation, and a balance between high performance and low energy consumption have become top issues to be solved in construction of green DCs.

Solution

CPO Technology

• This technology integrates the silicon optical module and switching chip onto one ASIC, greatly shortening the wiring distance between the switching chip and optical engine and reducing the SerDes drive capacity.

NPO and Cold-plate Liquid-cooling Technology

• This technology integrates the open optical engine interface (OPI) and switching chip onto the same PCB substrate to achieve chip decoupling with a standardized architecture, which significantly shortens the signal link.
• The cold-plate liquid-cooling technology adopts 3M fluorinated liquid (non-conductive) and independent cold plate for cooling. The built-in coolant micro leakage detection system eliminates the short circuit risk, ensuring high reliability of the rack.

Immersion Liquid-cooling Technology

• The switch is immersed in insulated and non-corrosive coolant. The heat is absorbed by the coolant and directly transmitted to the outside world through the cooling system, achieving heat dissipation and preventing the failure or aging of electronic components at high temperatures.