Huawei is known as a telecom equipment company or a consumer mobile devices company. But it’s all set to change soon. Huawei has ambitious plans to become a dominant player in the AI hardware and software segments.
With the sanctions imposed by US, Huawei wants to build its own stack of AI hardware and software without relying on external partners.
Eric Xu, Rotating Chairman announcing Ascend 910 AI Chip
The company has been working on a comprehensive, full-stack AI platform that compares to NVIDIA and Google offerings. Huawei claims that its stack is faster and efficient than the competition.
Let’s take a look at Huawei’s AI strategy.
Huawei AI Stack
There are only a few companies that own the combination of hardware and software layers for AI training and inference. Google does it with its TPU and TensorFlow stack and NVIDIA is popular for its integrated approach of GPU with CUDA and cuDNN.
Huawei is the most recent entrant into the space of integrated AI platforms. From the chipset to the high-level developer APIs, Huawei has built all the layers of an AI stack.
Ascend, Huawei’s family of AI chips offer acceleration for both model training and inference. It is the cornerstone of Huawei’s AI strategy delivering the horsepower required for dealing with complex neural networks.
The Ascend 910 chip that was announced recently belongs to the Ascend-Max segment that delivers the maximum power for training neural networks. For half-precision floating-point (FP16) operations, Ascend 910 delivers 256 TeraFLOPS. For integer precision calculations (INT8), it delivers 512 TeraFLOPS. Ascend 910's maximum power consumption is only 310W, which is much lower than the originally planned specs (350W).
Backed by years of experience in chip design and a deep understanding of customer scenarios, Huawei chose the unified Da Vinci architecture to develop the Ascend chips. Three unique, key technologies – scalable computing, scalable memory, and scalable interconnections – make this unified architecture ready for modern AI workloads.
Ascend-Nano chips are meant for accelerating neural network inferencing at the edge. They are similar to Google’s Edge TPU and NVIDIA’s Jetson Nano.
The CANN (Compute Architecture for Neural Networks) layer that sits above the chip layer acts as a software interface. It comes with a chip operator library and operator development toolkits. Delivering optimal development efficiency and operator performance, it meets the requirements of both academic research and enterprise scenarios.
CANN is similar to NVIDIA CUDA (Compute Unified Device Architecture) which is the software layer to talk to the underlying GPU cores.
The key component of CANN is its Tensor Engine, a highly automated operator development tool. It enables data scientists and developers to design complex deep learning networks involving intense mathematical calculations. Tensor Engine abstracts CANN through a programmer-friendly interface. It supports mainstream deep learning frameworks such as TensorFlow, PyTorch and PaddlePaddle.
Tensor Engine and its operators are Huawei’s equivalent of NVIDIA cuDNN, a library that makes CUDA accessible to AI developers.
MindSpore is Huawei’s own unified training/inference framework architected to be design-friendly, operations-friendly that’s adaptable to multiple scenarios. It includes core subsystems, such as a model library, graph compute, and tuning toolkit; a unified, distributed architecture for machine learning, deep learning, and reinforcement learning; a flexible program interface along with support for multiple languages.
MindSpore is highly optimized for Ascend chips. It takes advantage of the hardware innovations that went into the design of the AI chips.
Huawei’s MindSpore is an alternative to Google’s TensorFlow. Similar to the way Google optimized the combination of TensorFlow and TPU, Huawei has optimized MindSpore and Ascend. It’s also comparable to the combination of NVIDIA TensorRT and Jetson GPU family.
Apart from the hardware and deep learning frameworks, Huawei has also invested in an application enablement layer that acts as an ML Platform as a Service (PaaS) which is branded as ModelArts. It comes with integrated ML pipelines with all the stages such as data acquisition, data preparation, feature engineering, training, optimization and deployment built into the platform.
The PaaS layer also supports AutoML - a technique that bypasses the need to choose an algorithm - through a component called ExeML. Developers can consume in-built APIs for computer vision, natural language processing, and speech synthesis use cases.
Huawei claims that its application enablement layer, ModelArts, appeals to both beginners and experienced data scientists.
The PaaS layer in Huawei’s stack is comparable to Google’s AI Platform based on Kubeflow that promises a portable training and deployment environment. It also comes close to NVIDIA RAPIDS, a data prep framework that takes advantage of GPUs.
If Huawei manages to deliver on its vision, it is all set to become a dominant player in the AI hardware and software market.
