Authored by Mariana Rossi, cloud sales representative at Oracle, and Kellsey Ruppel, principal product marketing director at Oracle.
The authors want to thank Russell Anam, senior software engineer at iSIZE, for his contributions.
Figure 1: iSIZE offers deep learning for video delivery.
The need for access to high-performance GPUs for the development and training of AI models is growing, but many suppliers don’t offer cost-effective solutions. When developing AI solutions for video streaming, a clear demand exists for high-performance GPU instances and large volumes of storage and CPU and GPU memory to train a range of neural network models on batches of high-dimensional video data. Given the continuous development and testing that this entails, these features must always be available at a competitive price with high reliability levels.
iSIZE develops artificial intelligence (AI)-based software for optimizing video streaming. The company develops and licenses two products, BitSave and BitClear, and is developing advanced solutions for generative video content in a third stack of technologies called BitGen2D and 3D.
Goals for cloud migration
iSIZE BitSave is an AI-based perceptual preprocessing solution that allows conventional, third-party encoders to produce higher quality video at lower bitrate. Unlike traditional video processing methods, BitSave offers a cost-effective solution that can sustainably supply a seamless customer experience and enable companies to quickly adopt more efficient codecs.
BitClear is an AI-based neural network technology focused on denoising and upscaling heavily compressed video assets to recover original quality. BitClear enables customers to share, host, or otherwise provide higher quality media content that better engages the targeted audience. The solution achieves this goal by using pretrained AI models to remove compression artifacts and optionally upscale the video content.
The iSIZE BitGen2D and 3D stack of solutions is ongoing developments for generative neural network-based rendering within videoconferencing and telepresence and avatar technologies.
Training such advanced neural network-based solutions on video content requires advanced GPU and CPU capabilities and substantial memory to carry out effective training iterations. iSIZE explored multiple options for both inhouse and public cloud solutions and found Oracle Cloud Infrastructure (OCI) to be particularly appealing in comparison to all other alternatives. Specifically, the 24-7 availability of GPUs and CPUs with high reliability and competitive pricing for dedicated bare metal machines.
Suite of Oracle products used
iSIZE uses the following OCI products and services:
- OCI Compute service: Oracle offers secure and elastic compute options in the cloud that range from virtual machines (VMs) and bare metal servers at a lower cost. What sets Oracle apart from other cloud providers is the availability of bare metal dedicated machines with several GPU types at a competitive price.
- Identity and Access Management (IAM): IAM enables iSIZE staff to manage permission levels and access to resources. Compartments allow compartmentalization of permission and resources.
- Virtual cloud network (VCN): VCNs are customizable and private cloud networks on OCI. The VCN provides complete control over cloud networking environment like a traditional data center network.
- Cloud Guard: Cloud Guard provides a unified view of security postures across customer tenants. It allows customers to monitor, identify, achieve, and maintain a strong security posture.
- OCI Alarms: This service sends notifications across channels to operators when resources are at risk or may become unavailable.
- OCI Block Storage: This service can function like a typical hard drive, enabling you to create, attach, connect, move, and change volumes. The best thing about using Block Storage on OCI is the ability to adjust the speed of the block storage according to your need.
iSIZE did a 60-day POC with Oracle to test the superior price performance for A10 bare-metal against their other on-premises GPU provider. They chose A10 running on OCI due to availability, scalability, security, and superior price performance on bare-metal.
To deploy their AI training system, iSIZE first set up a subnet where all their bare metal instances reside. They manage subnets with a security list and a routing table. The security list keeps the platform secure by allowing limited specific traffic.
Figure 2: A graphic depicting an iSIZE Reference architecture diagram.
Then bare metal instances were started with the latest NVIDIA A10 GPU available. The instances consist of a mixture of container and custom instances. The platform is accessible on secured public web browser over SSL or directly by SSH with secure authentication. Oracle’s availability allows iSIZE to run continuously.
Since moving to OCI, iSIZE has been able to run 24-7 training with ample machines, which has allowed the company to reach and complete customer trials quickly. A single subnet with instance affinity has allowed the different instances to move and share data more quickly because an AI dataset can be up to 1 TB. This capacity has provided iSIZE with an expanded customer base. OCI’s superior price performance has allowed iSIZE to experience 22% in cost savings running their GPU’s on OCI vs. on-premises.
With the constant AI training and availability of more high-performance GPU during special loads, iSIZE has been able to respond to commercial needs quicker with newer and better models. Cloud Guard has also given the company another level of protection and peace of mind that its IP is protected.
The partnership between Oracle and iSIZE has allowed the company to adapt dynamically to customer demand and produce cutting edge AI models delivering higher quality video to end users.
With the help of Oracle customer support, who have been instrumental in helping the company at every step, iSIZE hopes to continue building on this success.
For more information on iSIZE and Oracle Cloud Infrastructure, see the following resources: