On February 26, Huawei announced that its Atlas 900 AI cluster has won the Global Mobile Awards 2020 (GLOMO Award) Future Technology Award. It is currently the world’s fastest AI training cluster, using Huawei’s self-developed 7nm Ascend 910. processor.
The GLOMO Awards Future Technology Award was set for the first time by the GSMA organizers to recognize technologies that transcend the times and reshape the world.
Huawei said that the Atlas 900 AI cluster won the GLOMO Future Technology Award this year, marking the industry’s high recognition of its technological innovation in the field of artificial intelligence.
Atlas 900 is currently the fastest AI training cluster in the world. It consists of 1024 Ascend 910 AI processors, with a total computing power of 256P~1024P FLOPS @FP16, equivalent to the computing power of 500,000 PCs.
It is understood that the Ascend 910 is currently the chip with the highest computing density on a single chip, and its computing power far exceeds that of Google and Nvidia. The Ascend 910 has a half-precision (FP16) computing power of 256TFLOPS, twice as much as NVIDIA’s Tesla V100, an integer precision (INT8) of 512TOPS, supports 128-channel full HD video decoding (H.264/265), and has a design power consumption of 350W ( The actual measurement reaches the specification and the computing power is only 310W).
In the model training benchmark of ResNet-50, the Atlas 900 broke the world record with a score of 59.8 seconds, and is the only product in the industry that can complete training in one minute.
It can be widely used in scientific research and business innovation, allowing researchers to train AI models such as images and videos faster, allowing humans to more efficiently explore the mysteries of the universe, predict the weather, explore for oil, and accelerate commercial processes such as autonomous driving.
The Atlas 900 AI cluster adopts three types of high-speed interconnection methods of “HCCS, PCIe 4.0, and 100G Ethernet”. Based on CloudEngine series data center switches, a 100 TB fully interconnected and non-blocking dedicated parameter synchronization network is built, which reduces network latency and reduces parameter synchronization latency by 10%. %~70%, making the AI model training process more efficient.
In terms of heat dissipation, Huawei has created an industry-leading heat dissipation system for the Atlas 900 AI cluster. It adopts a mixed liquid cooling solution, innovatively designs the industry’s strongest cabinet-level airtight insulation technology, and supports more than 95% of liquid cooling. A single cabinet supports ultra-high cooling power consumption of up to 50kW, achieving the ultimate data center energy efficiency with a PUE lower than 1.1.
At the same time, compared with the 8kW air-cooled cabinet, it can save 79% of the computer room space. The ultimate liquid cooling technology meets the needs of high-power, high-density equipment deployment, and low PUE, which greatly reduces the customer’s TCO.