In a bid to establish a foothold in an AI chip market that’s anticipated to be worth $91.18 billion by 2025, Huawei today brought to market the Ascend 910, a new chipset in its Ascend-Max family optimized for AI model training, and the Ascend 310, an Ascend-Mini series inferencing chip designed to tackle tasks like image analysis, optical character recognition, and object recognition. It also announced MindSpore, a computing framework intended to support AI app development.

Both chips were first detailed in October 2018, but this week marks their commercial debut. The Ascend 910 is aimed principally at datacenter workloads, while the Ascend 310 targets internet-connected devices like smartphones, smartwatches, and other internet of things (IoT) devices.

“Ascend 910 [in particular] performs much better than we expected,” said Huawei rotating chair Eric Xu at a press conference this morning in Shenzhen, China. “Without a doubt, it has more computing power than any other AI processor in the world.”

According to Huawei, the Ascend 910 delivers 256 teraflops of computing power for half-precision (FP16) floating-point operations and 512 TOPS (tera-operations per second) in integer precision (INT8) calculations while consuming about 310 watts of power. Huawei says that in a typical training session with ResNet-50, a popular image recognition benchmark, it’s about twice as fast as rival AI accelerator chips.

For the sake of comparison, Baidu’s Kunlun 818-300 AI chip, which was announced last July, can achieve 260 TOPS, while Amazon’s Inferentia machine learning processor offers scalable performance from 32 TOPS to 512 TOPS at INT8.

As for the Ascend 310, it packs a 16-channel FHD video decoder and is capable of 16 TOPS in integer precision (INT8) and 8 teraflops in half-precision (FP16). Huawei says it’s already being used by “leading” automakers in shuttle buses, new-energy vehicles, and autonomous driving, and it notes that its Ascend 310-touting Atlas series acceleration card and server are now a part of “dozens” of industry solutions developed by dozens of partners.

The Ascend 310’s performance puts it slightly behind startup Hailo’s recently announced Hailo-8, which is capable of 26 tera operations per second (TOPs) — more than double that of chips like Gyrfalcon’s Lightspeeur 2801, which maxes out at 9.3 TOPs, and CEVA’s NeuPro, which can reach 12.5 TOPs. And it’s leagues slower than Nvidia’s test chip, dubbed RC-18, which can hit 128 TOPs at 10 TOPs per watt.


Alongside the Ascend 910 and the Ascend 310, Huawei detailed MindSpore, an extensible end-to-end AI computing framework akin to Google’s TensorFlow and Facebook’s PyTorch. It scales across all devices, edge, and cloud environments, and it’s designed to be lightweight. MindSpore results in 20% fewer lines of core code than “leading” frameworks when dealing with a typical AI model for natural language processing, leading to a boost in efficiency of 50%, on average.

On the privacy side of the equation, Huawei says MindSpore doesn’t process data itself but instead taps model protection tech to ingest only gradient and model information that has already been processed. In this way, it ostensibly preserves sensitive data even in “cross-scenario” environments while ensuring that AI algorithms remain secure and trustworthy.

MindSpore will be made available in open source in the first quarter of 2020, with support for GPUs, CPUs, and other types of processors, in addition to Huawei’s Ascend chips. It’ll complement Huawei’s fully managed ModelArts platform, which provides full-pipeline model production services, including data collection and model development.

“Everything is moving forward according to plan. We promised a full-stack, all-scenario AI portfolio. And today we delivered,” said Xu. This launch is a new milestone in Huawei’s AI roadmap; it’s also a new beginning … We want to drive broader AI adoption and help developers do what they do best.”

The unveiling comes several weeks after the announcement of HarmonyOS, Huawei’s in-house operating system (OS) for mobile devices, in-vehicle systems, smart speakers, and wearables, and as the U.S. contemplates easing trade restrictions on the company and its subsidiaries as early as this year. But escalating tensions with China threaten to put the kibosh on those plans in the near term — Bloomberg reported yesterday that the White House will delay a decision about granting licenses that would allow U.S. companies (including Google and Intel) to restart business with Huawei, in light of China’s decision to halt purchases of U.S. farming goods.

Huawei founder and CEO Ren Zhengfei said in June that the U.S. ban would cost Huawei in the neighborhood of $30 billion in lost revenue.

The audio problem: Learn how new cloud-based API solutions are solving imperfect, frustrating audio in video conferences. Access here