Nvidia has unwrapped a new edge computing platform, the EGX range, which aims to make machine learning and AI viable as close as possible to the site of data consumption/production.
Edge computing is becoming increasing popular as enterprises look to lower latency and the efficient use of bandwidth. Nvidia wants to be the computational power behind the intelligence in devices out on the edge.
Speaking to investors during an earnings call, Nvidia CEO Jensen Huang noted that as AI systems grow and the data that they have to process expands to incredible scales, he believes that much of the processing of AI capabilities will happen in or close to the edge.
In that call he stated: “It’s too much data to move all the way to the cloud. You might have data sovereignty concerns, you want to have very, very low latency, maybe it needs to have multi-sensor fusion capabilities, so it understands the context better. For example, what it sees and what it hears has to be harmonious. And so, you need that kind of AI, those kind of sensor computing at the edge.”
In order to facilitate that intelligence Nvidia has created EGX: an optimised software stack operating the company’s own drivers integrated with Red Hat’s Openshift software and CUDA Kubernetes container capabilities.
The EGX platform is highly scalable, starting at a Jetson Nano moving all the way to Nvidia’s high performance T4 servers.
Bob Pette, VP of Enterprise and Edge Computing at Nvidia commented in a release that: “Enterprises demand more powerful computing at the edge to process their oceans of raw data — streaming in from countless interactions with customers and facilities — to make rapid, AI-enhanced decisions that can drive their business.”
“A scalable platform like NVIDIA EGX allows them to easily deploy systems to meet their needs on premises, in the cloud or both.”
Nvidia’s Edge Computing AI EGX Scaleable System
Nvidia’s Jetson Nano is a system on a module device, which comes in at just 70 x 45 mm. Within that small frame it packs Nvidia’s GPU Maxwell architecture and 128 CUDA cores. The CPU is a Quad-core ARM Cortex-A57 with a 6GB memory.
The Jetson Nano is capable of encoding and decoding 4K video at 30 frames. Nvidia says it can deliver 472 GFLOPs of performance for running AI algorithms and is capable of running multiple neural networks in parallel.
Due to the nature of edge devices, power consumption is a concern for any component in its delivery, yet the Jetson Nano is capable of running high performance task on just five to ten watts. Nvidia expects the device, as a standalone sale, to be available by June.
The Jetson device described above is at the low end of the scal: the top consists of a full server rack of Nvidia T4 servers, which are based on the company’s Turing Tensor Core architecture. Operating with 2,560 CUDA cores and 320 Turing Tensor cores the T4 is capable mixed-precision operations of 65 TFLOPS. The T4 has a GPU memory of 16 GB and is can process 300 GB per second.
Bruce King, senior principal data scientist at Seagate Technology commented in the release that: “At Seagate we have deployed an intelligent edge GPU-based vision solution in our manufacturing plants to inspect the quality of our hard disk read-and-write heads. The Nvidia EGX platform dramatically accelerates inference at the edge, allowing us to see subtle defects that human operators haven’t been able to see in the past. We expect to realize up to a 10 percent improvement in manufacturing throughput and up to 300 percent ROI from improved efficiency and better quality.”