Huawei Ascend 910 AI chip unveiled – the greatest computing density on a single chip

Huawei has announced the launch new ranges of processors aimed at providing huge amounts of computing power, optimised for artificial intelligence workloads.

Called Ascend 910 and Ascend 310, the new chipsets are based on an architecture called Da Vinci, and were unveiled at Huawei Connect 2018 in Shanghai.

Huawei’s rotating chairman Eric Xu also unveiled the company’s strategy and a full stack of products and services to realise the goals of the Da Vinci project.

In addition to the company’s new line of Ascend chips, it includes a new AI framework called MindSpore, a platform-as-a-service product called ModelArts, and a lower-level library called Compute Architecture for Neural Networks (CANN).

Huawei also showed off the Ascend-Max, a processor built on Huawei’s Da Vinci core using a 7-nanometer fabrication process, which has a maximum power draw of 350W.

Ascend-Max can achieve 256 TeraFLOPS, or 256 trillion floating-point operations per second, with half-precision (16-bit) floating-point numbers. With 8-bit integer numbers, it can achieve 512 trillion operations per second.

While the Ascend 910 chipset and Ascend-Max processors are designed to be deployed in cloud environments, the chips built on the Ascend 310 platform are designed for low-power computing.

The Ascend-Mini system-on-chip also uses the Da Vinci core, and is fabricated using a 12nm process.

It is capable of 8 TeraFLOPS using half-precision floating-point numbers, and 16 TeraOPS with 8-bit integers. Its maximum power draw is 8W.

Along with the Ascend-Mini, Huawei has also announced the Ascend-Lite, Ascend-Tiny, and Ascend-Nano – all based on the Ascend 310 chipset. These low-power chips are designed to be used in smartphones, smart accessories, and smart watches.

Huawei will use its Ascend 310-based chips in a new range of AI products:

  • Atlas 200 — AI Acceleration Module
  • Atlas 300 — AI Acceleration Card
  • Atlas 500 — AI Edge Station
  • Atlas 800 — AI Appliance
  • MDC 600 — Mobile Data Center

Xu said that Huawei will not sell its chips directly to third-parties, but will instead provide AI accelerator cards, modules, and servers to customers. These devices will start shipping in February 2019.

The table below summarises the applications that Huawei’s new Ascend 910 and Ascend 310 chips are designed for.

Chip Device Edge Cloud
Ascend chip Nano Tiny Lite Mini Mini Multi-Mini Max
Application Earphones Always-on / Wearable Smartphone Laptop Industrial PC Edge Server Data Centre
Compute power 20 MOPS 100 GOPS 1-10 TOPS 10-20 TOPS 10-20 TOPS 10-100 TOPS 200+ TOPS
Power budget 1mW 10mW 1-2W 3-10W 3-10W 10-100W 200W+
Model size 10KB 100KB 10MB 10-100MB 10-100MB 100MB+ 300MB+
Latency <10ms ~10ms 10-100ms 10-500ms 10-500ms ms~s ms~s
Inference Yes Yes Yes Yes Yes Yes Yes
Training No No Yes Yes Yes Yes Yes

Now read: Huawei and DTPS stress the importance of building an ICT talent ecosystem

Latest news

Partner Content

Show comments


Share this article
Huawei Ascend 910 AI chip unveiled – the greatest computing density on a single chip