World of Huawei Ascend: Future with NPUs

Kubilay Tuna
Huawei Developers
Published in
7 min readFeb 9, 2023
Mr. Eric Xu introduces Ascend 910 Chip at HAI 2020 Conference

Introduction

Have you ever heard about NPUs? Now, you will meet the true power itself!

Huawei released the full-stack Ascend AI products at the Huawei Ascend Innovation (HAI) 2020 conference, setting a new milestone as Huawei and its partners are making strides forward to bring AI to reality. Huawei making quick entry into the field of AI offers a competitive AI computing infrastructure by giving the message that nothing will be the same from now on with the solutions it offers. Huawei’s AI computing solutions boast advantages in superior computing power, openness, ease of use, security, and low consumption. This revolutionary transformation helps us design the future.

Ascend AI, pure amazing power itself!

Let’s meet the future more!

Ascend Computing Industry is a full-stack AI computing infrastructure based on Ascend AI products and software. Ascend includes Ascend AI Processors, series hardware, CANN (Compute Architecture for Neural Networks), and AI computing frameworks, as well as development tool-chain and management and O&M tools. Briefly, the purpose of creating an AI environment is to support every industry possible by bringing the future to everywhere possible.

Supporting Industries by Ascend

So, why ascend?

These AI solutions have been widely used in sectors including healthcare, transportation, finance, university education, scientific research, astronomical exploration, smart city, and oil exploration. But why prefer Huawei Ascend?

The answer is very obvious!

There are many benefits from Ascend Computing Industry. Specifically, some of these are having an open ecosystem enriched with its partners, affordable prices that keep you within your budget, and open, collaborative, and transparent AI built with personal privacy and data security scopes. That makes all system reliable and allow you to make development according to your needs.

No one wants their private data to share, right? Let’s keep it safe with Huawei.

What are the differences between NPU and GPU? Why would I prefer NPU?

Let’s take a closer look at the benefits of NPU!

NPU stands for Neural Processing Unit, and GPU stands for Graphics Processing Unit. Both NPUs and GPUs are specialized hardware designed to improve the performance of certain types of computations, such as those used in machine learning and other complex computational tasks. One key difference between NPUs and GPUs is that NPUs are specifically designed to accelerate the computations used in artificial neural networks, which are the basis of many machine learning algorithms. In contrast, GPUs are designed to accelerate the calculations used in graphics rendering, but can also be used for other types of parallel computations. Another key difference between NPUs and GPUs is the type of calculations they are optimized for. NPUs are optimized for the half-precision (FP16), highly parallel calculations used in artificial neural networks, while GPUs are optimized for full-precision (FP32) calculations used in graphics rendering and other tasks. This means that NPUs can perform certain types of computations more efficiently than GPUs, but GPUs may be better suited for other types of calculations.

In summary, NPUs and GPUs are both specialized hardware designed to improve the performance of certain types of computations, but they are optimized for different types of calculations and have different strengths and weaknesses. If we explain everything in short, NPU is a microprocessor that accelerates machine learning tasks to the moon 🚀. NPUs are created with a different architecture than GPUs. You can do video editing, gaming, machine learning, and simultaneously many tasks together. But NPUs focus only on Machine Learning tasks. So that that focus creates nearly x10 faster training and inference performances compared to GPUs.

Do you want to hear the true power of AI computing? So, great news! Right addresses 🙌

Huawei: Ascend 910, The World’s Most Powerful AI Processor

The limits of this incredible power are not over yet, Ascend Industry provides even more. Extreme performance is the most desired thing developers want. With Atlas 900 cluster, the total computing power reaches 256P~1024P FLOPS FP16, which is equivalent to the computing power of 500,000 high-performance PCs. This is the world’s fastest training cluster and the industry’s fastest ResNet-50@ImageNet performance.

Is it just the NPU that provides all this power?

No way! As you guess, powerful hardware needs powerful software. Thus Huawei developed a full-stack solution including hardware and software stack.

Let’s dive into full-series hardware! Before doing that, the best thing to do is to become familiar with some concepts such as training and inference.

In the context of machine learning, training, and inference refer to two different phases of the machine learning process.

>> Training, is the phase of the machine learning process in which an algorithm is presented with a large dataset and uses that data to learn the underlying patterns and relationships. During training, the algorithm adjusts its internal parameters to minimize the error between its predictions and the known correct answers in the training data.

>> Inference, refers to the phase of the machine learning process in which the trained algorithm is used to make predictions on new, unseen data. Inference typically happens after the training phase, and the algorithm uses the knowledge it has learned during training to make predictions on new data.

In summary, training is the phase of the machine learning process in which the algorithm learns from data, while inference is the phase in which the trained algorithm is used to make predictions on new data.

For these two different types of solutions, Huawei provides two different products in the most efficient way. One called Ascend 310 for inference and one called Ascend 910 for training. Huawei created two separate products because both have different purposes and advantages. For instance, if you need only inference on your production line. You do not have to pay for training in AI Units. Only an inference card will be enough to bring you the power of AI. If you want to train huge models with big data, you can use 8 Ascend 910 Training NPU in parallel to build with tremendous power. That’s why Huawei created flexibility for Ascend.

Huawei provides wider range of power to provide future down to smallest detail.

Let’s be real, different tasks require different solutions. If everyone needs everything in one, then everyone would have eaten their supper with Swiss Pocketknife but the reality is not 🤭

So, how about energy consumption, carbon emission, and carbon footprint?

We protect our future and nature 🌎 So we also prepared an environment-friendly product. For example, a liquid-cooled computer room saves air conditioning, PUE< 1.1 .Reduce Data Center Carbon Emissions. Environment-friendly products with future technology create a bright future. Energy consumption in total is way lower than Huawei’s competitors. We care about our children’s future and our planet's future.

Huawei HPC Liquid Cooling Solution

After a little but important additional information, let’s get back to our main topic and continue to take a closer look at the Huawei Ascend full-series hardware.

Atlas Full-Series Hardware

Power of Open-Source

As you know, the usage and power of open-source is increasingly widespread. Huawei uses open source for almost every product to expand knowledge and the future everywhere.

There are several open-source platforms for Huawei Ascend Development. Let’s explore together!

  • HiAscend: You can find every documentation, and tutorial on Huawei Ascend's main page.
  • AscendHub: Like dockerhub, you can use your Ascend devices inside prebuild docker with minimum effort (ready to use).
  • Gitee Ascend Page: You can find any tutorial and examples about Ascend computing.
  • Ascend Guideline: You can find every product in detail with guidelines.

Conclusion

In this article, we have learned about NPU, which is a term Huawei introduced to the literature, and made an introduction to Ascend Full-Stack infrastructure. Also, we briefly explained that Huawei supports the open source community and green energy while providing these strong infrastructures by giving some useful links. We’ll delve deeper into Huawei Ascend Full-Stack Solutions in our next articles with hands-on.

Congratulations on following this article till the end!

Congrats :)
Stay Tuned for Our New Articles!

“Sometimes life is going to hit you in the head with a brick. Don’t lose faith!” -Steve Jobs

References

--

--

Kubilay Tuna
Huawei Developers

Machine Learning Engineer || Passionate about machine learning, data science, and programming