The SwiftStack Data Platform for AI is a transformational announcement for the Accelerated Computing market and builds upon our “World’s 1st Multi-cloud AI/ML Data Management Solution” announcement at NVIDIA’s GPU Technology Conference in March earlier this year.
Accelerated Computing Market Transitions
Consider the transitions that have happened in this market even in the small period since March 2019: Businesses are adopting deep learning data pipelines beyond computer vision—including for Conversational AI (also known as natural language processing, NLP, and natural language understanding, NLU) and anomaly detection for a variety of use cases.
NVIDIA released the T4 GPU for inferencing—optimized for power and with Tensor-RT support. The “Smart Everything Revolution” has begun with autonomous IoT devices based on 5G Edge networks. At the Mobile World Congress just last week, NVIDIA released the EGX supercomputer at the edge working with Telco partners like Erricson for Telco—especially vRAN (Virtual Radio Access Network) use cases.
Transfer learning is gaining popularity for several verticals where customers want an easy path to adopt AI/ML/DL. NVIDIA has released SDKs like Clara for Healthcare, Metropolis for Smart Cities, and NVIDIA Jarvis for Multi-modal AI (vision, speech and sensors) to enable transfer learning. Deep neural nets like BERT and project Megatron are significantly bigger than Resnet and VGG and need training for several parameters, which increases the need for performance exponentially and is driving the increasing use of NVIDIA DGX-2 based SuperPODs.
SwiftStack AI Architecture—Built on The SwiftStack Data Platform
As NVIDIA Partner Network Solution Advisors, we at SwiftStack—with the help of our customers and partners—were able to foresee some of these market transitions, which became the basis of new capabilities in the SwiftStack Data Platform and the SwiftStack Architecture for AI. The SwiftStack AI Architecture is targeted to address the following use cases:
- Autonomous Vehicles
- Healthcare and Life Sciences
- Humanitarian Assistance and Disaster Response
- Intelligence, Surveillance, and Reconnaissance (ISR)
- Smart Cities
- Cyber Security
- Telco VRANs (Virtualized Radio Access Networks)
The SwiftStack Data Platform
The SwiftStack Data Platform is comprised of the following building blocks:
- SwiftStack Data Storage and Management Platform—including SwiftStack’s core storage technology and SwiftStack 1space.
- Distributed Data Access Layer—including filesystem and object access methods in the core, ProxyFS Edge, and third-party ecosystem partnerships. The new ProxyFS Edge capabilities provide file services for edge devices and edge aggregation points, optimizing scale-out file access from edge to core.
- SwiftStack Client—including end-user data access, dataset management, and metadata indexing and search.
The SwiftStack AI Architecture enhances the following data pipeline stages:
- Data Collection—at the Edge, Core and Cloud along with annotation, labeling, ETL processing, etc.
- Workspaces—for CI/CD workflows
- Deep Neural Network (DNN) Training—for the fastest time to insights
- DNN Validation—simulating real-world scenarios to test the trained DNNs
- Inferencing—on the edge and core
- Lifecycle Management—with versioning, tracing and reproducibility
Edge to Core to Cloud Engineered Solution
The SwiftStack AI Architecture delivers a customer-proven edge-to-core-to-cloud engineered solution with best-of-breed reference architectures, production-grade testing with 500 GPUs, performance numbers and sizing guidelines. We feel confident that this architecture blueprint will provide answers to IT challenges facing CIOs and data engineers in the modernization of IT infrastructure to support the demanding needs of deep learning workflows.
You can learn more about the SwiftStack AI Architecture by reading the newly published white paper.