Graphcore And Spell Collaborate For End-To-End ML Operations

by | Feb 25, 2022 | In the News

Companies also announce free IPU Test Drive with Graphcore AI hardware and Spell software.

Building, training, optimizing, and deploying AI models at scale can be excruciatingly difficult. Especially for large models that need to be optimized to run well at scale and need iterative hyper-parameter tuning to achieve desired accuracy. Since large models are so important to Graphcore, the company is now partnering with Spell to provide an end-to-end platform to support and automate the development process.

Who is Spell?

Spell helps address the challenges for AI developers with MLOps designed for the deep learning development pipeline. Spell operationalizes deep learning for natural language processing, computer vision and speech recognition at scale, bringing together popular tools and frameworks in a single software platform. Management tools are critical for these large networks of accelerated systems. Note that NVIDIA recently purchased Bright Computing to address this need for HPC clusters. Spell is a more focussed platform tailored for ML operations support, helping with initial model training and automating the optimization process, as well as integrating with deployment tools and cloud service providers. The IPU hardware has been seamlessly integrated with the Spell orchestrator, letting developers streamline their end-to-end MLOps workflow through Spell’s command-line interface.

Developers can now automate pipelines for training, optimization, testing, and inference for Graphcore IPUs, and Spell takes care of all infrastructure configuration behind the scenes. Spell also offers a model registry to let you track versioning of different models and monitor model performance.

Graphcore’s IPU Pod systems currently support up to 256 Intelligence Processing Units. Graphcore

Graphcore and Spell launch free IPU test drive

Providers of very large AI platforms such as Graphcore, SambaNova, and Cerebras struggle to provide access to potential developers and make the experience with early access a success. Now armed with Spell, Graphcore can confidently let developers experience the hardware on their own. Using browser-based notebook access to IPUs in the cloud via the Spell platform, developers can get 6 hours of free access to road-test IPU systems with a range of popular AI models, including natural language models such as BERT, computer vision models including EfficientNet, Graph Neural Networks (GNNs) like TGN. Now developers can quickly get familiar with key programming principles and experience the ease of use of the Graphcore software stack, which has matured significantly over the last year.

Conclusions

As we have noted in earlier articles, the Graphcore software ecosystem continues to grow rapidly, especially now that the community is embracing and contributing to the open source code the company has released. Adding a platform for MLOps is the natural next step, and it looks like they picked a good partner with Spell. Finally we applaud the Free IPU Test Drive; more AI startups should do something like this, but few have the confidence in their software platform to create a satiusfying self-guided experience!