Huawei today announced the Ascend AI full-stack software platform, which includes CANN 3.0, a heterogeneous computing architecture, MindStudio, a full-process development toolchain, and MindX, an Ascend application enablement, covering basic software to application enablement.
Huawei said at the Ascend AI New Product Global Launch (HAI 2020) in Shenzhen that its AI computing solutions have been widely used in healthcare, transportation, finance, higher education, scientific research, astronomical exploration, smart cities, oil exploration and other fields, based on their advantages in super computing power, openness and ease of use, security and low consumption.
CANN 3.0
Infrastructure, as the only way to connect hardware and software, is a crucial point to leverage the computing power of AI.
At the conference, the new "3.0" version of CANN, first released in 2018, completes the upgrade of the unified heterogeneous computing architecture.
CANN 3.0 supports backward compatibility and evolution to create a compute base for end-side cloud-wide scenario collaboration. Users only need one set of application code to be compatible with 10+ device types, 14+ operating systems, and multiple AI frameworks, enabling full-scenario AI development.
CANN 3.0 provides two ways to develop operators that are both efficient and powerful.
CANN 3.0 unlocks the full performance of AI hardware. Currently, the CANN 3.0 high performance operator library has 1000+ hardware affinity operators with deep optimization.
At the same time, CANN 3.0 supports extreme performance through deep graph optimization, automatic graph splitting and fusion, intelligent data pipeline optimization, and other technologies.
Based on the highly adaptive hardware and software combination, Huawei Atlas hardware with CANN 3.0 achieves industry-leading performance in mainstream inference and training models.
In video scenarios, the Atlas 300I can simultaneously process 80 channels of 1080p, 25FPS HD video, which is twice the number of simultaneous channels processed in the industry.
Huawei Announces Ascend AI Full Stack Software Platform: CANN 3.0, MindStudio, MindX
MindStudio 2.0
At the conference, MindStudio 2.0, a full-flow development toolchain, was unveiled.
MindStudio provides developers with an easy-to-use set of development tools to efficiently complete end-to-end development, allowing developers to "carry the whole process" from operator development, model training, model reasoning, application development, and application deployment without having to work on different tools.
MindStudio is able to support training and reasoning scenarios. During training, the Ascend training acceleration tool uses its unique Less BN (intelligent identification of unnecessary BN operators in the network) and random freeze algorithm to significantly improve the efficiency of model training.
In inference scenarios, the Ascend model compression tool accelerates the inference process with its unique intelligent algorithms.
MindX
MindX 1.0, which debuted at the HAI conference, helps integrate AI computing into every industry through "2+1+X". 2" is the deep learning platform MindX DL, "1" is the preferred model library ModelZoo, and "X" is the intelligent edge platform MindX Edge. " is to enable SDKs across industries.
The MindX DL deep learning platform focuses on unified management and scheduling of data center devices and computing resources, enabling partners to rapidly develop commercial deep learning systems.
The advantage of MindX Edge is that it enables one-click deployment of cloud-side collaborative inference services across multiple devices, with a platform memory overhead of only 256MB and CPU utilization of only 3%.
ModelZoo, a library of preferred models by scenario and framework, is the core driving force behind the two deep learning platforms, and currently offers more than 20 high-performance models by scenario and framework.
According to the conference, this number will increase to more than 50 in the second half of the year.
Huawei Ascend builds industry SDKs that encapsulate Huawei's basic algorithms and accumulated experience in the industry AI domain.
At present, Huawei Songshan Lake's Atlas-based smart manufacturing solution has been implemented to manufacture "Atlas" with "Atlas" in Huawei's production line of computing products, allowing AI to run through every aspect of manufacturing.
The 80+ production lines in Huawei's factories have already been applied to cover servers, 5G, terminals and other products. The experience accumulated from these large-scale applications will be encapsulated in the manufacturing SDK.
The Vision + Manufacturing SDK will be officially launched in October 2020.
