Huawei and DeepSeek launch CloudMatrix 384 supernode: a new benchmark for AI computing power
Recently, Huawei and DeepSeek jointly released their major productsCloudMatrix 384 Super Node, marking the entry of AI computing power infrastructure into a new era. This breakthrough technology integrates Huawei's hardware advantages and DeepSeek's algorithm innovation capabilities, providing super computing power support for large-scale AI training.
1. Core performance parameters
index | parameter |
---|---|
Computing unit | 384 Ascend 910B processor |
Computing power scale | 147.5 PFLOPS (FP16) |
Internet bandwidth | Huawei's self-developed EDGE interconnect architecture, with bandwidth up to 768GB/s |
Energy efficiency ratio | 1.2TFLOPS/W |
Supports model size | Exceeding trillion-dollar parameter big model training |
2. Highlights of technological breakthroughs
1.Heterogeneous computing architecture: Through the deep integration of Huawei Atlas 900 PoD and DeepSeek distributed training framework, the utilization rate of computing resources is increased by 40%.
2.Dynamic elastic scheduling: Support real-time expansion and reduction of training tasks, shortening the failure recovery time to seconds
3.Green energy-saving design: Use liquid-cooled heat dissipation technology, with PUE value as low as 1.08
3. Industry application prospects
Application areas | Typical cases |
---|---|
Autonomous driving | Supports parallel processing of simulation training data at 10 million kilometers |
Drug development | Molecular dynamics simulation efficiency is 300 times |
Financial Technology | It can realize the iteration of high-frequency trading strategies at millisecond level |
AIGC | Support weekly training of 50 billion parameter multimodal large model |
4. Comparison of market products
product | Computing Power (PFLOPS) | Energy efficiency ratio | Node Scale |
---|---|---|---|
CloudMatrix 384 | 147.5 | 1.2TFLOPS/W | 384 nodes |
NVIDIA DGX H100 | 132 | 0.9TFLOPS/W | Node 256 |
Google TPU v4 | 120 | 1.1TFLOPS/W | 2048 chip |
5. Progress in ecological construction
Currently, more than 20 leading companies have joined the CloudMatrix ecosystem plan, including:
- Academic institutions: Institute of Automation, Chinese Academy of Sciences, Institute of Intelligent Industry, Tsinghua University
- Cloud service providers: Alibaba Cloud, Tencent Cloud, Volcano Engine
- Industry users: Xiaopeng Motors, WuXi AppTec, Ping An
6. Expert opinion
The Secretary-General of the China Artificial Intelligence Industry Development Alliance said: "The launch of CloudMatrix 384 fills the gap in the field of domestic AI computing power infrastructure in the field of super-large-scale clusters, and its innovative heterogeneous collaborative architecture provides a new paradigm for the industry."
According to official reports, the first batch of super nodes will be deployed in Shanghai Lingang computing power hub and Gui'an Huawei Cloud Data Center in Q2 2024, and the annual production capacity is expected to reach 50 super node clusters.
check the details
check the details