Movidius targets low power for AI at the edge

Technology News |
By Nick Flaherty

The 16nm Myriad X processor has a power envelope of 1.5W and was developed at the company design centres in Dublin, Ireland and Romania.

The Neural Compute Engine is a purpose-built hardware accelerator designed to dramatically increase performance of deep neural networks without compromising the low power characteristics of the Myriad processor family. It uses an array of MAC blocks and directly interfaces with the intelligent memory fabric to handle 16bit floating point and 8bit fixed point convolutional neural net (CNN) operations with performance of 1TOPS.

“We’re on the cusp of computer vision and deep learning becoming standard requirements for the billions of devices surrounding us every day,” said Remi El-Ouazzane, vice president and general manager of Movidius which has been part of Intel’s New Technology Group for the last year. “Enabling devices with humanlike visual intelligence represents the next leap forward in computing. With Myriad X, we are redefining what a VPU means when it comes to delivering as much AI and vision compute power possible, all within the unique energy and thermal constraints of modern untethered devices.”

The NCE can be linked to 16 proprietary programmable 128bit VLIW vector processor cores to provide ten times the performance of the previous generation of processor at 4TOPS.

The Myriad X also has 16 configurable MIPI lanes for camera data, supporting up to 8 HD resolution RGB sensors. A stereo depth accelerator block can simultaneously process 6 camera inputs (3 stereo pairs) each running 720p resolution at 60 Hz frame rate.

The high-throughput inline ISP ensures streams are processed at high speeds, while new hardware encoders provide support for 4K resolutions at both 30 Hz (H.264/H.265) and 60 Hz (M/ JPEG) frame rates. Other featured interfaces include USB 3.1 and PCI-E Gen 3.

The chip has 2.5 Mbytes of memory with 400 GByte/s internal bandwidth to reduce the power consumption from having to go off chip, but there are is a package option for external memory (the MA2085) and one with 4Gbits of low power LPDDR memory in the package (the MA2485).



Linked Articles
eeNews Power