MindSpore

What MindSpore is

MindSpore is an AI computing framework from Huawei. A framework of this kind supplies the building blocks a developer assembles into a neural network — layers, operators, the machinery for computing gradients, and the plumbing for running a job across many machines — so that a team does not have to write that infrastructure itself.

MindSpore's stated design goal is to cover all scenarios natively: the same framework is meant to serve models that run on a device, at the edge and in the cloud, rather than requiring a separate toolchain for each setting. It is co-optimized with Huawei's Ascend AI processors, and it also supports GPU and CPU.^[1]

Where it came from

Huawei first detailed MindSpore in August 2019, alongside the Ascend 910 training chip and the Ascend 310 inference chip.^[2] That timing put the framework in place as the software counterpart to Huawei's own silicon: one chip aimed at training, one at inference, and a framework tuned to run on them.

Work on the public code had already begun before the formal announcement. The project's repository, mindspore-ai/mindspore on GitHub, was created on 11 February 2020 and is released under the Apache-2.0 license.^[3]

On 28 March 2020, at HDC.Cloud in Shenzhen, Huawei announced that MindSpore had officially gone open source on Gitee.^[1] The announcement was made by Chen Lei, Huawei's chief scientist for MindSpore.^[1]

Main features

Huawei's pitch centred on how much code a developer has to write. The company said MindSpore uses about 20% fewer lines of code than leading frameworks for typical natural language processing models, and that this improves development efficiency by more than 50%.^[2] These are Huawei's own figures for its own framework, not an independent measurement.

MindSpore launched with two companion modules. MindInsight handles training visualization and debugging — letting a developer watch how a training run is progressing rather than inferring it from a single number at the end. MindArmour deals with adversarial robustness and model security, the problem of inputs deliberately perturbed so that a trained model misclassifies them, and of keeping a deployed model from being attacked or copied.^[2]

Who built it

Huawei said MindSpore was developed with external collaborators including Imperial College London, Peking University and the University of Edinburgh.^[2] The involvement of university groups is a reminder that a framework is partly a research artifact: the techniques it packages, such as methods for making models robust to perturbed inputs, come out of academic work as much as out of product engineering.

Where it sits

MindSpore belongs to a class of projects in which a hardware company builds the software layer above its own chips. The Ascend processors provide the compute; MindSpore provides the programming model that developers use to put that compute to work; and the open-source release on Gitee, under a permissive licence, is what lets people outside Huawei read the code, use it and modify it.

That combination — silicon, framework and an open repository — is the sense in which MindSpore matters beyond one company's product line. It made Huawei one of the few organizations offering both the chips and the framework for training and serving neural networks, and it gave developers who work with Ascend hardware a native way to write their models instead of adapting an outside framework to the hardware themselves.