MENU

Movidius upgrades VPU with on-chip neural compute

Movidius upgrades VPU with on-chip neural compute

Technology News |
By Peter Clarke



Although Movidius has not said what manufacturing process the Myriad X is implemented in it does claim that it is capable of delivering an aggregate performance of over 4 trillion operations per second (TOPS) and that this makes the chip sufficiently power efficient to perform vision and artificial intelligence applications within drones, smart cameras, smart home, security, VR/AR headsets, and 360 cameras. The neural compute engine is a capable of 1 TOPS of compute performance on deep neural network inferences.

The neural compute engine comes with 16 SHAVE vector processor cores the MyriadX includes additional upgrades to imaging and vision engines including additional programmable SHAVE cores, upgraded and expanded vision accelerators, and a 4K ISP pipeline with support for up to 8 HD sensors connecting directly to the VPU.

As with Myriad 2, the Myriad X VPU is programmable via the Myriad Development Kit (MDK) which includes development tools, frameworks and APIs to implement custom vision, imaging and deep neural network workloads on the chip.

The Myriad 2 VPU was capable of running neural networks and indeed it was this capability that led to its introduction of the Fathom Neural Compute Stick, based on its Myriad 2 vision processor in 2016 (see Movidius shows neural network stick). The Myriad X VPU extends this capability so that it can reach 10X the performance for applications that require multiple neural networks running simultaneously, the company said.

Next: More than 20 hardware accelerators


There are more than 20 hardware accelerators on chip to perform tasks such as optical flow and stereo depth without introducing additional compute overhead. And the imaging and vision hardware accelerators and the Neural Compute Engine, can be operated alongside an array of 16 C-programmable VLIW vector processors, all accessing a 2.5Mbytes of common on-chip memory.

The chip comes in two variants in a package measuring 8.7mm by 8.5mm. The MA2085 interfaces to external memory while the MA2485 comes with a 4Gbit LPDDR4 memory in-package.

Related links and articles:

www.movidius.com

News articles:

Intel buys Movidius to ramp computer vision

Movidius shows neural network stick

Google’s deep learning comes to Movidius

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News

Share:

Linked Articles
10s