Meet Synthesia
Accelerating Large Language Model Training in the Cloud
Synthesia was founded with a mission to empower everyone to make professional quality video content – without cameras, microphones, or studios. Using AI, Synthesia is here to radically change the content creation process and unleash human creativity for good.
Build Lifelike Video Content With Generative AI
Synthesia provides text-to-video and text-to-speech capabilities that are intuitive to use and simple to get started with, providing AI-generated avatars to create lifelike video content without requiring camera equipment or human actors. Synthesia provides an intuitive web interface to craft compelling video content. Organizations can match avatars with voices, display graphics, and apply animations to create engaging and polished videos.
Artificial intelligence research is central to Synthesia’s service. Synthesia’s AI researchers continue to push the boundaries of what’s possible, constantly training their models to generate AI avatars that produce more realistic renditions of human movement and speech. New model versions will provide even more lifelike avatars, while continuous model tuning expands the catalog of avatars and graphics content available to Synthesia customers. Synthesia researchers constantly push to improve their models through continued model training cycles and new data sets to support this effort.
“WEKA helped us with engineering peace of mind. It saved us a massive deployment effort that would have delayed bringing our product to market.”
The Challenge
In the text-to-video space, fast time to market is critical to success, with newer, more powerful models constantly coming to market. From day one, AI research experienced delays due to manual data operations and performance limitations in the legacy Lustre storage system.
Low GPU Utilization
Poor performance and scale limitations associated with their initial legacy cloud-based Lustre system created bottlenecks in feeding model data into the GPU cluster, resulting in idle GPU resources.
Manual DataOps
Prior to WEKA, the infrastructure team spent hours manually copying the correct data sets from Amazon S3 directly into the training cluster, wasting many hours and introducing data consistency challenges that required constant infrastructure monitoring during model training epochs.
AI Model Training Delays
Low GPU infrastructure utilization, combined with manual data operations slowed down AI training cycles and resulted in slow time to market for new features.
“During AI model training, WEKA can fully saturate our GPUs, helping us get more performance out of our AI infrastructure and scale when needed.”
The Solution
Synthesia relies on the WEKA Data Platform to drive high-performance data operations for the entire AI model training and tuning environment. WEKA’s software runs on a cluster of Amazon EC2 i3en instances; the aggregate NVMe flash attached to the instances – now at 80 TB with the ability to linearly scale with developer needs – provides the high-performance storage layer that delivers “local” performance over a shared file system directly in AWS. The WEKA environment extends to hundreds of TBs of object storage in Amazon S3 for massive capacity at a low cost.
“WEKA gives us increased flexibility that simply wasn’t possible with any other solution we tried.”
Outcomes with WEKA
The WEKA Data Platform has helped the Synthesia team avoid manual data management processes, simplifying data operations and reducing costs. WEKA’s zero-copy, zero-tuning architecture eliminates the need for infrastructure teams to manually copy data from Amazon S3 into the training cluster and maintains data consistency across every node. Synthesia is experiencing improved researcher productivity thanks to the quantum leap in performance.
Simplified Data Operations
Synthesia eliminated manual data operations by relying on WEKA zero-copy, zero-tuning data architecture.
Improved GPU Utilization
Synthesia relies on WEKA to saturate their GPU infrastructure, driving faster AI model training.
Increased Productivity
Synthesia is experiencing improved researcher productivity thanks to the quantum leap in performance provided by WEKA.
“With WEKA, we can migrate our entire data set virtually with the push of a button. We continued doing model training across two AWS Regions simultaneously over two weeks while we migrated the compute.”
Cloud Migrations Simplified
Synthesia recently migrated AWS Regions to support continued growth in their GPU-accelerated compute cluster. During the migration, the Synthesia infrastructure team took advantage of the flexibility built into WEKA to make the move simple and easy from a data perspective. Synthesia was able to rapidly stand up an entire new WEKA environment in the new region thanks to WEKA deployment automation via Terraform. Next, the Synthesia team leveraged industry-unique WEKA Snap to Object capability for the data migration. Snap to Object creates a fully usable snapshot of production data – including all data and metadata – and saves that snapshot to an object store in Amazon S3. Synthesia could generate a fully usable copy of their AI training data in just a few minutes and move it from one region to another. Synthesia could run AI model training side by side in both regions to continue running experiments while migrating the compute stack. Once ready, the shift to the new region was seamless, enabling a zero downtime migration between AWS Regions.
How Synthesia Maximizes GPU Efficiency and Accelerates AI Model Training with WEKA on AWS
Learn more about how Synthesia increases researcher productivity with WEKA
Dive a little deeper
