UbiOps helps Bayer Crop Science to scale with computer vision workloads across GPUs rapidly and easily. The project will create a new milestone providing a flexible service, making it possible to run unpredictable loads at high throughput and low cost.
The collaboration between UbiOps and Bayer Crop Science started in the beginning of 2022. Dr. Alexander Roth, Head of Engineering – Digital Crop Protection at Bayer explains: “We were looking for a solution that adapts to our computer vision model, inferencing workloads on-demand. Whilst at the same time minimizing or potentially avoiding an upfront investment in IT infrastructure with GPU resources. One of the biggest challenges of existing solutions is to deliver inferencing results reliably in real-time, given unpredictable customer behaviour. With UbiOps we have found a solution that is able to cope with this challenge by scaling on-demand across GPUs rapidly.”
Computer vision models from Bayer Crop Science
The Crop Protection Innovation Lab at Bayer Crop Science – part of the German multinational pharmaceutical and life sciences company Bayer AG – uses artificial intelligence to shape future digital farming by driving the current digital transformation in agriculture. Such AI-based solutions are crucial as they help improve harvest quality and accuracy – known as precision agriculture. For example, thousands of farmers worldwide use Bayer’s solutions for disease detection, weed classification, pest control, and more. With the MagicScout app from Bayer Crop Science, you can perform lighting-fast detection of weeds, even without internet in the field. Moreover, you can detect leaf diseases in the most important crops within seconds.
Unpredictable AI workloads at scale
Dr. Roth: “Once we bring our computer vision models into production, we do not know upfront how they will be used, i.e., how their workloads will scale over time. Agriculture is a seasonal business that is heavily influenced by environmental conditions, e.g., weather, pests, etc. Hence, model loads can significantly vary per day or even per minute. As a result, an underlying infrastructure that can optimize load across GPUs rapidly to cope with unpredictability and still provide high throughput is essential. That’s how we discovered UbiOps.”
UbiOps is a serving layer that can adapt to unpredictable workloads by rapidly scaling across multiple GPUs. The auto-scaling algorithm, scale-to-zero functionality, and dynamic distributed resource pooling help data analytics teams to instantly scale across GPUs to process dynamic workloads in time, and at the same time minimize GPU costs.
UbiOps focuses on the ease of the deployment process. Deployment can be a daunting task for a data scientist. However, UbiOps makes it effortless to go to production, thanks to the beginner-friendly deployment format. There’s no need to learn a new technology to deploy. Plus, there’s additional support in place to facilitate data scientists in this process.
The architecture diagram shows that a user captures an image using MagicScout app (Mobile Client). The image is uploaded to AWS cloud where it goes through a series of AWS services that perform authentication, validation, and routing before reaching UbiOps. Deployments powered by serverless GPUs on the UbiOps perform AI inference and service user requests. Once the request is serviced, the response goes through the architecture back to MagicScout, where the user sees the results on their smartphone.
Dr. Roth: “UbiOps not only offered us a technical solution to this challenging problem, but their pay-as-you-go model enables us to be cost-effective. With UbiOps there is no need for an upfront investment in IT infrastructure, which saves us from under-/over-provisioning. With GPUs being a scarce and expensive resource, using them effectively when scaling up is very important. The on-demand offering of UbiOps ensures that there’s GPU availability, with the option to scale very rapidly. Also, with UbiOps’ scale-to-zero functionality we don’t need to pay for GPU resources if the application is not being used, e.g., off -season”.
Reliable inferencing and scalable GPU training
So far, the collaboration with UbiOps has been fruitful. Dr. Roth: “After testing UbiOps for the first time, our AI scientists were positive that the anticipated goals will be achieved. The reliability of the platform is very high, which is very important to us. The platform has 99.99% uptime, and the speed of request handling is particularly consistent. Currently, we are further scaling up the number of applications. I believe if we can profit from it, so can other organizations that require on-demand rapid scaling across GPUs with high reliability.”
Reliable inferencing is an essential step to make computer vision ubiquitous and bring value to customers. However, computer vision models are erroneous, i.e., they provide results based on high confidence. This is a huge challenge in agriculture as every decision made can have a severe impact on yield. As a result, Bayer`s goal is to constantly improve computer vision models in collaboration with customer feedback from the field, i.e., customer feedback needs to be incorporated in a constant retraining loop.
In the continuing collaboration between the Crop Protection Innovation Lab at Bayer and UbiOps, the goal is to address the additional issues in such a retraining loop. Frequent retraining relies on ensured availability of GPUs, which can be guaranteed by pooling distributed GPU resources dynamically. For that UbiOps will support Bayer to provide reliable and scalable GPU training that is available on-demand.