Deep Learning Now 199 Times Faster on CPU, Thanks to MagicNet From PQ Labs

January 15, 2019

FREMONT, Calif.--(BUSINESS WIRE)--Jan 15, 2019--A Lego car equipped with a low power 0.9GHz processor and a tiny camera is able to do all the self-driving car tasks in the wild, chasing and playing with your cats and dogs, recording and uploading videos, collecting fruits, following your expeditions - a scene that may only happen in a Sci-Fi movie, will soon become in reality thanks to PQ Labs MagicNet, a new deep learning network that accelerates AI 199 times faster on CPU.

Running deep learning tasks without a GPU or AI chips hardware can actually be faster? That seems to be mission impossible for decades until January 2019 when PQ Labs demonstrated a jaw-dropping deep learning solution to surprise every visitor in the CES tech show.

By embedding a 0.9GHz Intel processor into a Lego car, the toy acquires Artificial Intelligence skills instantly to runs self-driving AI tasks such as detecting objects or obstacles. In the past, the toy would have to carry a heavy computer case with a high graphics card installed in order to have such AI computing power.

MagicNet doesn’t stop there. It is not only for small and low latency neural networks running on low power devices. When equipped with an Intel i7 processor, the performance scales up well making it run even faster than NVIDIA GPU card on deep learning tasks such as object detection: MagicNet running at an astonishing 718 fps on Intel i7 v.s. Tiny Yolo running 292 fps on NVIDIA TITAN X or 1080Ti graphics card, both achieving the same accuracy conditions.

It is the first time in the history that Intel processor wins over NVIDIA GPU by a large margin in the deep learning field. NVIDIA has easily dominated this market for many years with virtually no competition, thanks to the NVIDIA’s proprietary CUDA, CUDNN deep learning library, and hardware design.

MagicNet is still evolving. While most neural networks running on GPUs focus on low resolutions such as 224x224, 300x300, 608x608, 1280x720, etc. MagicNet Face 4K network is able to run at 4096x2048 resolution. The Magic Convolution operation of 4096x2048 in the neural network’s first layer helps to extract very fine and detailed features that no other network has tried before. Thus, MagicNet is able to detect 2,000+ human faces in one shot from just one single image, while other AI product in the market can only detect 200 faces at maximum.

Speed is not the limit. Contrary to many neural network optimization schemes, MagicNet has not based the speed boost on network quantization yet. The next MagicNet version will come with int8 quantization support, which will make MagicNet run even faster and more energy efficient.

For more information about MagicNet, visit the website http://www.pqlabas.ai and signup for Demos, Evaluation Kit and latest updates.

View source version on businesswire.com:https://www.businesswire.com/news/home/20190115005408/en/





SOURCE: PQ Labs Inc.

Copyright Business Wire 2019.

PUB: 01/15/2019 10:00 AM/DISC: 01/15/2019 10:01 AM


Update hourly