Huawei steps up its AI game with the Ascend 910 AI processor launch

• Ascend 910 processor “the most powerful AI processor”
• Won’t be sold as standalone, packaged with other products, cloud services

At a lake on Huawei’s Bantian Base in Shenzhen, China, one may be able to find black swans gracefully skimming across the serene surface. 

Not quite something you find on the headquarters of one of the world’s largest technology companies, but there is symbolic meaning behind the black swan. As a Huawei representative tells us, the black swan signify “unpredictability” – they serve as a reminder that the unexpected happens, and the company needs to prepare for it.

The waters of the company are certainly tumultuous now, as Huawei continues to face pressure from the US government. It is on a blacklist which restricts American companies from doing business with it, key of them being companies like Google and Qualcomm, which Huawei relies on for key technology.

Yet the company put on a brave face, appearing unfazed as it launched the Ascend 910, an AI processor they are claiming to be the most powerful. With this, Huawei is pitting itself against major American giants like Qualcomm and NVIDIA – at the same time, defying the continued US pressure by showing that they can still develop core technology.

Growing the AI brain

Artificial Intelligence isn’t a completely new venture for the company. Huawei’s previous products, which include the Ascend Nano and Tiny Chips, as well as its AI-embedded Kirin chips, have mostly brought AI to devices like smartphones and gadgets.

With the Ascend 910, however, it is diving headlong into the market for data centre and cloud-based processing – in line with the Chinese government’s goal of being a worldwide leader in AI.

“We have been making steady progress since we announced our AI strategy in October last year,” says Huawei’s Rotating Chairman Eric Xu. “Everything is moving forward according to plan, from R&D to product launch. We promised a full-stack, all-scenario AI portfolio, and today we delivered. This also marks a new stage in Huawei’s AI strategy.”

Announced last year, the Ascend 910 belongs to Huawei’s series of Ascend-Max chipsets. It is, as the company explains, used for AI model training. Companies using AI applications require huge amounts of data to train smart algorithms, typically taking days or weeks.

Huawei, however, claims that its chip can process more data faster than its competitors. In a typical training session based on ResNet-50, the combination of Ascend 910 and their newly-announced MindSpore AI computing framework performs two times faster at training AI models compared to other mainstream training cards using TensorFlow.

The company also says that test results show that the processor delivers on its performance goals “with much lower power consumption than originally planned.”

In terms of performance specifics, the Ascend 910 delivers 256 TeraFLOPS for half-precision floating point (FP16) operations. For integer precision calculations (INT8), it delivers 512 TeraOPS. Max power consumption, the company notes, is only 310W, which is lower than its initially-planned specs of 350W.

Spreading the MindSpore

MindSpore, launched alongside Ascend 910, is an AI computing framework that supports development for AI applications. These frameworks are critical to making AI application development easier.

Huawei says that MindSpore marks significant progress towards their three development goals for  AI framework, which it announced in 2018. The three goals are Easy Development, Efficient Execution and Adaptable to All Scenarios – MindSpore is stated to be able to readily adapt to different deployment needs.

Adding to that, MindSpore is said to help ensure user privacy, as it only deals with gradient and model information that has already been processed. It doesn’t processed the data itself, Huawei notes, so private user data can effectively be protected even in cross-scenario environments. MindSpore also has its own built-in model protection technology.

MindSpore uses an “AI Algorithm As Code” design concept, which allows developers to develop advanced AI applications with ease and train their models more quickly.

How quickly? According to Huawei, in a typical neural network for natural language processing (NLP), MindSpore has 20% fewer lines of core code than leading frameworks on the market. It also helps developers raise their efficiency by at least 50%.

“MindSpore will go open source in the first quarter of 2020,” Xu announces during the launch. “We want to drive broader AI adoption and help developers do what they do best.”

Going forward

The fascinating aspect about the Ascend 910 is that it won’t be offered as a commodity chip to other companies. Instead, Huawei will use it to bolster its own data centre services. With MindSpore, which is optimised for Huawei’s own chips, Huawei is positioning itself not as a component supplier but as a systems house.

“The Ascend series of processors will not be provided to the market as a standalone product, but embedded in our computing platforms, acceleration card and cloud services,” Xu says.

With both the chips and framework, the company provides a more compelling reason for customers to adopt the full suite of Huawei tech.

During the press conference, Huawei demurred when asked for the pricing of the Ascend 910, but  did say it would be competitive with higher-end devices in the market. Mind, those from NVIDIA currently run as much as US$10,000 each.

The company also did not reveal its Ascend 910 rollout plan, so there’s no certainty of how it will be offered to those interested in SEA region. When asked about what the company’s AI strategy is for India, Xu says that it “will be no different from our overall strategy.”

The strategy here is that they will be developing the Atlas computing platforms based on the Ascend ecosystem – products that can be provided to universities, partners and enterprises to develop AI applications and address industry-specific issues, Xu elaborates.

“Additionally, we have training and cloud services based on the Ascend processors to be made available to enterprises.”

The company also announces that the Ascend 310 processor – the company’s first commercial AI System on a Chip (SoC) in the Ascend Mini series – has seen wide adoption in a broad range of products and cloud services since its launch.

Concluding the launch event, Xu says that Huawei will debut more AI products at its upcoming conference, Huawei Connect 2019, which is set to happen between September 18 and 20 in Shanghai. Watch this space!

Related stories:

Malaysia Airports collaborates with Huawei Technologies Malaysia in Airports 4.0

Celcom deploys Malaysia’s first 5G live cluster field trial

Micron’s ‘fab’ jobs for the new fab plant in Singapore