Recently, Stanford University published the latest DAWN Deep Learning Benchmark (DAWNBench). Alibaba Cloud broke the records set by Google and other enterprises, taking first place in four categories in the image recognition competition. This proves that Alibaba Cloud can provide the world’s fastest artificial intelligence (AI) computing services.
DAWNBench is one of the most authoritative competitions in the AI field, comprehensively measuring the strength of solutions in terms of deep learning optimization strategies, model architectures, software frameworks, clouds, and hardware. In Image Classification on ImageNet, Alibaba Cloud took first place in four categories: training time, training cost, inference latency, and inference cost, breaking the records held by Google and other companies.
(Alibaba Cloud ranked first in training time, training cost, inference latency, and inference cost)
According to DAWNBench, it takes only 2 minutes and 38 seconds to train ImageNet with 1.28 million images when using Alibaba Cloud’s heterogeneous computing service, and only 0.0739 milliseconds to identify an image when using the Hanguang 800-based AI service. Alibaba Cloud also set a world record in training cost and inference cost.
Alibaba Cloud’s record-setting performance can be attributed to its Apsara AI Acceleration (AIACC) and Hanguang 800 products. AIACC is the industry’s first unified AI acceleration engine for mainstream deep learning frameworks such as TensorFlow, PyTorch, MXNet, and Caffe. Working with the same hardware platform, AIACC can significantly improve the performance of AI training and inference.
Hanguang 800 is Alibaba’s first in-house chip. It is the world’s most powerful AI inference chip, with unrivaled performance and energy efficiency. The Hanguang 800-based AI cloud service can implement 1,600 trillion deep learning computations every second.
In recent years, deep learning models have become more and more complex and they require significantly more computing resources and time. The high cost of AI computing power limits the adoption of AI applications. Therefore, improving AI performance and reducing the cost have become the top priority for enterprises.
The Alibaba Cloud computing platform has greatly improved AI performance through hardware and software collaboration and innovation. At present, our AI computing services are extensively used in many AI scenarios, such as image recognition, speech recognition, content recommendation, autonomous driving, and natural language understanding. For example, Horizon Robotics, an AI startup in China, uses Alibaba Cloud AI computing services for deep learning training, tripling their performance while significantly reducing training costs.
Alibaba Cloud provides an open-source AIACC-based AI solution and will launch Hanguang 800-based AI cloud services in the near future.