A host of startups at the GPU Technology Conference in Israel this week are showing off the extraordinary acceleration gains and efficiency the NVIDIA DGX Station provides for their deep-learning work.
Companies like Cognata — which creates technology for testing autonomous vehicles in VR and won last year’s $100,000 Inception award at GTC Israel — praise DGX Station, the world’s first personal AI supercomputer, for accelerating their work up to 10x.
Another startup, TRACXPOiNT, which is creating an AI-infused shopping cart, recouped the cost of purchasing its DGX Station within two months, based on what it saved in paying for using GPU instances in the cloud.
“It’s no exaggeration to say that we depend on our DGX Station every day, sometimes every hour of every day,” said Danny Atsmon, CEO of Cognata. “It’s an enormous advantage in speeding up our work and getting to market fast.”
While it provides the computing power of over 400 CPUs, it uses only about one-twentieth as much power. It creates only one-tenth the noise of a workstation — about as quiet as a typical office ventilation system.
An integrated system purpose-built for AI, DGX Station comes with fully optimized hardware and software. This enables companies to get started in just an hour, compared with potentially a month of setup time required to build similar systems.
DGX Station’s deep learning and analytics performance is unmatched, providing:
- 72x the deep-learning performance of CPU servers
- more than 100x the speedup for analyzing large datasets versus a 20-node Spark server cluster
- full versatility for both deep learning training and inferencing at over 30,000 images a second.
Here’s how four of Israel’s hottest young companies use DGX Station.
It’s estimated that an autonomous car would need 11 billion miles of test drives to achieve the same level of accuracy as a human driver. Cognata puts that within reach by using state-of-the-art deep learning simulation to test vehicles in virtual reality through computer-generated landscapes, complete with other cars, pedestrians, buildings and varying weather conditions.
By using a DGX Station, Cognata shaved off years of training time, accelerating its efforts by 10x beyond what it could achieve even in a GPU-powered workstation. The system enables its team to simultaneously run dozens of training jobs. So Cognata can rack up millions of virtual miles on which it can finetune autonomous responses.
AnyVision, which has grown to more than 170 employees in just four years, applies proprietary technology and convolutional neural networks to provide face, body and object detection. It enables capabilities such as facial recognition for ticketless entry to sporting events and visual identification for two-factor authentication for banking applications.
DGX Station enables AnyVision to train 8x faster than on a sophisticated GPU-powered workstation, while detecting individual identities against a database of 115 million faces in 200 milliseconds.
This Jerusalem-based startup monitors the effects of mutations and drugs on cancer patients, enabling oncologists to provide precision, personalized treatments.
Using DGX Station, NovellusDX was able to train 4x faster by eliminating the need for large data transfers to the cloud and save $70,000 on an annual basis, for an eight-month payback period.
NovellusDX also improved its accuracy by a factor of 10x in using its own deep learning framework to quantify the level of intra-cellular signaling pathway activity from millions of images of mutating cells.
This startup is bringing the convenience of online shopping to retail with an AI-powered shopping cart that visually recognizes items in stores, communicates with suppliers to get real-time offers, and enables shoppers to pay digitally for their cart’s contents without stopping to scan items at the checkout counter.
TRACXPOiNT’s cart is fully integrated with hardware and GPU-accelerated software. Training on a DGX Station provided a 3x performance increase, and the company made its money back after just two months of 24/7 training versus GPU-accelerated cloud solutions. It conducts inferencing using TensorRT software running on the NVIDIA Jetson embedded platform, which enables it to recognize up to 100,000 different products in under a second.