Huawei unveils the full stack Ascend AI software at HAI 2020

Huawei released the full-stack Ascend AI software at the Huawei Ascend Innovation (HAI) 2020 conference, setting a new milestone as Huawei and partners are making strides forward to bring AI to reality. The release includes the Compute Architecture for Neural Networks (CANN) 3.0, MindStudio full-pipeline development toolchain, and MindX for Ascend application enablement. The software portfolio covers all the way from developing basic software to deploying AI in the real world.

Tony Xu, President of Huawei Ascend Computing Business, introduces the goal of this release: To deliver AI computing power of the ultimate accessibility, ease of use, and performance.

Tony Xu, President of Huawei Ascend Computing Business

CANN 3.0
Basic architecture connects hardware and software and is crucial to unlocking AI computing power. Huawei released CANN 1.0 at HUAWEI CONNECT 2018. CANN 3.0 signifies an upgrade to a unified heterogeneous computing architecture.

CANN 3.0 supports backward compatibility. Using CANN as the computing foundation for collaboration across devices, the edge, and the cloud, enterprises can deploy one set of code on over 10 device form factors, more than 14 operating systems, and multiple AI frameworks. This enables AI development in all scenarios. In addition to being inclusive, CANN 3.0 is simplified. AscendCL provides a unified programming interface to meet developers’ requirements for custom frameworks and operators, enabling efficient development. CANN 3.0 provides two operator development modes: TBE-DSL and TBE-TIK, to deliver both efficiency and performance.

CANN 3.0 fully releases the performance of AI hardware by leveraging competitive features. These features include a high-performance algorithm library composed of over 1000 operators deeply optimized for Ascend; in-depth graph optimization, automatic graph splitting and fusion, and intelligent optimization of data pipelines.

Based on the highly adaptive software and hardware, Huawei Atlas hardware equipped with CANN 3.0 surpasses industry peers on mainstream inference and training models. In video scenarios, the Atlas 300I inference card, powered by Ascend chips, can process 80 channels of 1080p 25FPS HD videos concurrently, which is twice that of its industry counterparts.

MindStudio 2.0
MindStudio 2.0 is released to facilitate development from end to end, covering operator development, model training, inference, application development, and application deployment. Developers gain unprecedented ease in AI development and no longer need to perform such operations on different tools.

MindStudio supports training and inference scenarios. It provides an Ascend training acceleration tool, which uses a unique Less BN algorithm to intelligently identify unnecessary BN operators in the network, and a random freezing algorithm, to greatly improve the model training efficiency. For inference scenarios, the Ascend model compression tool uses unique intelligent algorithms to accelerate the inference process.

MindX: 2+1+X Help Deliver AI Computing to Industries
Regarding AI computing, in addition to performance, Huawei attaches as much importance to its ease of use. At HAI 2020, Huawei released MindX 1.0, which features 2+1+X to accelerate the convergence of AI and industries. 2 indicates the MindX DL deep learning platform and MindX Edge intelligent edge platform; 1 indicates the ModelZoo pre-trained model library; X indicates the software development kits (SDKs) for various industries.

MindX DL focuses on unified management and scheduling of data center devices and computing resources, enabling partners to quickly develop deep learning systems.

The advantage of MindX Edge is that it can deploy edge-cloud collaborative inference services on various devices at one click. The platform requires only a memory overhead of 256 MB and a CPU usage of 3%.

ModelZoo is classified by scenarios and frameworks. It is the core driving force for the two deep learning platforms: MindX DL and MindX Edge. Currently, Huawei has provided more than 20 high-performance models in ModelZoo. The number is expected to exceed 50 in the second half of this year.

Huawei Ascend takes one step further to enable AI applications across industries by providing industry SDKs that are based on Huawei’s basic algorithms and expertise in AI.

More and more partners are using Huawei Atlas to empower the digital transformation of industries. The instant translator system iFLYREC, provided by iFLYTEK, translates for HAI 2020 online, is based on the Atlas AI solution.

According to Tony Xu, Huawei’s full-stack Ascend AI software helps maximize the performance of AI hardware and lower the entry barriers to AI application. Ascend serves as the platform for AI infrastructure and also the enabler of AI for industries. Huawei is working with ecosystem partners to build a vibrant Ascend computing ecosystem and together ascend to pervasive intelligence.