Huawei has introduced its new Atlas 960E SuperPoD system, which is the first in the world to utilize Near-Packaged Optics (NPO) technology. This innovation aims to enhance computing capabilities and speed up the training and inference of large models. According to David Wang, Vice Chairman of the Board and Rotating Chairman of Huawei, the company's strategy focuses on maximizing the value of devices by developing SuperPoD and SuperCluster systems.
The launch of the Atlas 960E SuperPoD comes as basic models approach 10 trillion parameters, leading to increased infrastructure requirements. Currently, the highest capacity achieved by scientists is 2 parameters. The Atlas 960E relies on UnifiedBus and Hi-ONE technologies, featuring an advanced orthogonal architecture and full liquid cooling. UnifiedBus unifies memory addresses and connects computing resources in a single environment.
The Atlas 960E SuperPoD system can expand to include 4096 NPU units, providing a computing power of up to 8 EFLOPS at FP8 precision and 16 EFLOPS at FP4 precision. By integrating 5500 Hi-ONE units, the system eliminates the need for 48,000 traditional 800G optical communication units, reducing power consumption by over 550 kilowatts and achieving a 99.8% availability rate.
Huawei's Hi-ONE is the first commercially produced optical communication engine using NPO technology, with a transmission capacity of 7.2 terabits per second. It is the only product in the sector that adopts NPO with an integrated light source. This technology enables the Atlas 960E to efficiently handle demanding AI workloads.
In addition to the Atlas 960E, Huawei has applied its SuperPoD architecture to multiple systems, including the TaiShan 950 SuperPoD. This system supports 4096 computing nodes and 256 terabytes of unified memory, enhancing intelligent agent environments. Huawei has also launched the OceanStor M900 storage system, providing petabyte-level KV Cache memory for L3.5 layers.
Huawei's Atlas 960E SuperPoD systems can be interconnected to build a SuperCluster with up to 512,000 NPU units using the Clos architecture, or up to 1 million NPU units with the Multi-Rail architecture. This scalability enables the creation of powerful AI computing infrastructures that can handle complex tasks.
The introduction of the Atlas 960E SuperPoD with NPO technology marks a significant milestone in the development of AI computing infrastructure. By providing high-performance computing capabilities and efficient data transmission, Huawei's innovative solution is poised to accelerate AI training and inference, driving advancements in various industries and applications.
Key points
- Huawei launches Atlas 960E SuperPoD with NPO technology for AI training.
- The system provides up to 16 EFLOPS at FP4 precision and reduces power consumption.
- Huawei's Hi-ONE is the first commercially produced optical communication engine using NPO technology.
