Movidius: A stick for neural-networks from Intel for $79.
After Intel took over the machine-vision specialist company Movidius last year, a low-cost USB stick with hardware for deep learning is now on offer from the company for only $ 79. The Movidius Neural Compute Stick contains an SoC offering computing power of up to 100 Gigaflops which simply plugs into a commercial PC’s USB 3.0 port...
After Intel took over the machine-vision specialist company Movidius last year, a low-cost USB stick with hardware for deep learning is now on offer from the company for only $ 79. The Movidius Neural Compute Stick contains an SoC offering computing power of up to 100 Gigaflops which simply plugs into a commercial PC’s USB 3.0 port.
The USB stick contains a Myriad-2-SoC with twelve VLIW 128-bit vector processors. The core runs at 0.9 V with a 600 MHz clock. A fast 2 MB on-chip memory gives a transfer rate of 400 GB/s. FP16 as well as FP32 and integer operations with 8, 16 and 32 bit accuracy are supported. The chip uses 28-nm technology and is divided into 20 power islands, that can be individually powered on and off depending on processing load to optimize power management and guarantee high efficiency.
The stick is especially well suited to applications employing simulated neural networks and ideal for implementing tasks such as object or face recognition in security camera applications. Movidius claims the stick can natively run neural networks built using the Caffe framework. The stick will be useful for adding more power locally when training and designing new neural nets. According to Movidius its also possible to use multiple sticks, giving linear performance boost with each added stick. Priced at $79, the Movidius Neural Compute Stick is available in the USA.
The USB stick contains a Myriad-2-SoC with twelve VLIW 128-bit vector processors. The core runs at 0.9 V with a 600 MHz clock. A fast 2 MB on-chip memory gives a transfer rate of 400 GB/s. FP16 as well as FP32 and integer operations with 8, 16 and 32 bit accuracy are supported. The chip uses 28-nm technology and is divided into 20 power islands, that can be individually powered on and off depending on processing load to optimize power management and guarantee high efficiency.
The stick is especially well suited to applications employing simulated neural networks and ideal for implementing tasks such as object or face recognition in security camera applications. Movidius claims the stick can natively run neural networks built using the Caffe framework. The stick will be useful for adding more power locally when training and designing new neural nets. According to Movidius its also possible to use multiple sticks, giving linear performance boost with each added stick. Priced at $79, the Movidius Neural Compute Stick is available in the USA.
