No NN-dedicated hardware on Qualcomm IoT processors

No NN-dedicated hardware on Qualcomm IoT processors

Technology News |
Qualcomm has announced two 10nm processors in support of IoT applications computing at the edge of the network, the quad-CPU QCS603 and the octa-CPU QCS605.
By Peter Clarke

Share:

The chips are described as being part of the Qualcomm Vision Intelligence Platform although they provide support for both visual and audio processing. The SoCs include advanced camera processing software, machine learning and computer vision software development kits (SDKs), and connectivity and security technologies.

Both processors provide a lot of CPU, GPU, DSP and other processing resources. However, as with Qualcomm’s Snapdragon line of processors it appears that while Qualcomm is prepared to support machine learning in software it is yet to provide dedicated hardware to accelerate this function. Both chips claim to support Tensorflow, Caffe and Caffe2 and other machine learning development environments using the Snapdragon Neural Processing Engine.

An engine sounds like hardware, right? Well yes but it this case it is a programming interface. As in the case of Snapdragon back in 2016 for now the neural network software piggybacks on the existing Kryo CPU, Adreno GPU and Hexagon DSP cores inside the SoC. Back then it was the Snapdragon 820 processor (see Qualcomm offers neural network SDK for Snapdragon processor). Now it is the same processing resources inside the QSC603 and QSC605.

It is notable that many IP core providers and fabless chip startup companies are now offering dedicated machine learning hardware. ARM, an IP provider with which Qualcomm has long-standing cooperation, has launched two machine learning  processors for which the IP is due to become available in mid 2018 (see ARM launches two machine learning processors).

Next: What’s the difference?


The key difference between the two IoT processors is in terms of CPU resources and performance.

Both are based on a big-little configuration. The QSC603 offers two 64-bit Kryo300 Gold cores and two Kryo300 Silver cores and these operate at a clock frequency of up to 1.6GHz and 1.7GHz respectively. It is strange that the big processor has the lower maximum clock frequency. The QSC605 provides two Kryo300 Gold cores running at up to 2.7GHz and two Kryo300 Silver cores running at 1.7GHz.

Apart from that the resources are the same on both: the Adreno 615 GPU, the Hexagon 685 vector processor, the Spectra 270 image signal processor as well as DSPs for sensors and audio. With these components comes support for up to 2×2 802.11ac Wi-Fi with MU-MIMO, Bluetooth 5.1. Audio software includes the 3D audio suite, Aqstic and aptX audio.

While the neural processing engine offers optimization and debug and is compatible with Tensorflow, Caffe and Caffe2 frameworks, Open Neural Network Exchange interchange format, Android Neural Networks API, and the Qualcomm Hexagon Neural Network library, the fact that the hardware is not optimized for machine learning does leave something to be desired.

The Qualcomm QCS603 and QCS605 SoCs are currently sampling, Qualcomm said but did not give an indication of how much the components would cost.

Related links and article:

www.qualcomm.com

News articles:

Qualcomm offers neural network SDK for Snapdragon processor

Ceva goes non-DSP with neural processor

Eta raises funds for spiking neural networks

Qualcomm-backed startup announces AI processor family

€5 octacore MCU covers analytics, fusion at the edge

ARM launches two machine learning processors

ST preps second neural network IC

Linked Articles
eeNews Analog
10s