Ascend 910
AI training chip with DaVinci cores and HBM2 memory
The Ascend 910 is an artificial intelligence training processor designed by Huawei. It uses thirty-two DaVinci cores and high-bandwidth memory to deliver high compute performance for deep learning workloads.
- Contains 32 DaVinci AI cores for training
- Uses four HBM2 memory stacks
- Operates at a 350 W power envelope
- Co-packaged with Nimbus V3 I/O chip
- Supports CCIX and RoCE networking
What it is
The Ascend 910 is a flagship artificial intelligence training chip from Huawei. It was presented at Hot Chips 31 in 2019 as a competitor to NVIDIA training solutions. The processor relies on a custom architecture called DaVinci. It combines thirty-two of these AI cores with high-speed memory to handle large-scale deep learning tasks. The chip is designed to operate within a 350 watt power envelope. Huawei positions it as part of a broader ecosystem that scales from edge devices to large server clusters.
How it came about
Huawei developed the Ascend 910 to reduce dependence on external AI hardware suppliers. The company showcased the chip in 2019, highlighting its ability to scale across different market segments. The design aims to provide a complete alternative to existing GPU-based training platforms. Huawei intended to sell millions of devices based on the DaVinci core architecture. This effort supports their domestic market and global infrastructure ambitions.
Inside the chip
The chip features thirty-two DaVinci AI cores arranged in a mesh network. This interconnect has six rows and four columns. It includes four stacks of HBM2 memory for high bandwidth data access. A separate Nimbus V3 chip handles input and output operations. This I/O die is co-packaged with the main processor and memory. Two dummy dies are included for mechanical stability. The design also integrates video decoding engines and supports CCIX connectivity. Huawei claims significant performance in half-precision floating point calculations.
How it was made
The physical layout includes the main Virtuvian die and the Nimbus V3 I/O chip. Four HBM2 memory stacks surround the core logic. Two additional dummy dies, each 110 square millimetres, ensure structural integrity. The total area including these dummy components is reported as 1028 square millimetres. The package integrates these elements into a single module for server installation.
Where it went
This processor is used in AI training servers and large computing clusters. Huawei builds systems with eight Ascend 910 modules per board. These boards often use water cooling to manage heat. The servers typically pair the AI chips with dual Intel Xeon Scalable processors. The architecture supports scaling up to large clusters similar to competing supercomputing platforms. It is intended for data centres requiring high-throughput machine learning capabilities.
Versions and family
The Ascend family includes other chips for different purposes. The Ascend 310 is designed for AI inference at the edge and uses Arm A55 cores. A mid-range Ascend 610 is also mentioned, offering higher bandwidth per core but fewer cores than the 910. The DaVinci core itself is designed to scale from internet of things devices to automotive applications and large training units.
Why it matters
The Ascend 910 represents Huawei's strategy to create an independent AI hardware ecosystem. It challenges the dominance of NVIDIA in the AI training market. By offering a scalable solution from edge to cloud, Huawei aims to capture a significant share of the domestic market. The chip supports standard interconnects like RoCE networking, facilitating integration into existing data centre infrastructures.
Questions people ask about Ascend 910
What is the DaVinci core?
The DaVinci core is Huawei's custom processing unit designed for artificial intelligence tasks. It serves as the building block for the Ascend chip family, scaling from small edge devices to large training processors.
How much memory does the Ascend 910 have?
The chip integrates four stacks of HBM2 memory. This high-bandwidth memory allows for fast data access, which is critical for training large deep learning models efficiently.
What is the power consumption of this chip?
The Ascend 910 is designed to run at a power envelope of 350 watts. This high power draw necessitates advanced cooling solutions, such as water cooling, in server deployments.
Can it be used for AI inference?
While the Ascend 910 is focused on training, the DaVinci architecture also powers the Ascend 310 chip. The Ascend 310 is specifically designed for AI inference tasks at the edge.
Good reading and sources
There is no English Wikipedia article on Ascend 910 yet, so the facts on this page come from the sources listed above: servethehome.com/huawei-ascend-910-provides-a-nvidia-ai-training-alter, hiascend.com/en/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.