As AI is invading more and more our daily lives, a new type of hardware is making its way: the AI-powered processors, that is to say, the processors specially designed to run Artificial Intelligence algorithms.
Some of the readers may recall the Hollywood movie named ‘Small Soldiers’ (1998 ). In the movie, a factory put by mistake new revolutionary military AI-powered processors into a group of several toys aimed to be sold to kids.
The result is that the toys are behaving fully autonomously and create – of course – panic and chaos everywhere.
While in the context of 1988 such a scenario was clearly anticipation, in 2020 (so 30 years later) this is not at all unrealistic because we are seeing – among other things – the rise of A.I. powered processors everywhere.
This is just the beginning but as always with disruptive technologies, it will just grow exponentially years after years or even months after months.
Here we wish to present to the reader a quick guide to the latest AI processors available for sale as of October 2020.
Table of Contents
GPU or Graphical Power Unit, while originally targeting video applications, has become a de facto standard in AI hardware, and companies traditionally providing cards for the videogame market such as Nvidia have now turned to AI hardware development.
This is a high-end card which targets researchers in AI as well as designers of AI systems. It can be scaled and used for a lot of applications such as market pricing prediction or data analysis in general.
The NVIDIA Titan RTX is powered by a specific architecture, the Turing architecture. The Turing microarchitecture features among other large matrix operations for AI as well as Deep Learning Super Sampling (DLSS).
Other similar microarchitectures include the Volta and Ampere Architectures.
The specifications of the card are truly impressive:
The card allows processing huge datasets such as the ones typically met in science applications. It is fully integrated with a set of data science libraries named the RAPIDS suite and using CUDA-X AI SDK.
The typical retail price of a Titan RTX card is USD 2,500- USD 3,000.
There are more varieties of AI-powered GPU cards provided by NVIDIA. Other AI products are also available like the NVIDIA QUADRO GV10.
Radeon instinct is the GPU card for Deep learning manufactured by AMD. Note that AMD also produces AI-powered microprocessors for PCs.
The brand is divided into several products: MI-6, MI-8, MI-25, and MI-50.
Here are the specifications of the MI-25 card:
AMD provides software such as MIOpen which offers support for the biggest AI frameworks (Tensorflow, Caffe, Theano, etc)
The typical retail price of an MI25 Radeon card is around USD18,000
An important application of A.I. processors is for the Internet Of Things (IoT) devices.
Indeed these devices often need image recognition, and classification automatic decisions for example and they deeply depend on deep learning methods, especially Convolutional Neural Networks (CNNs).
GAP8 is a French revolutionary processor provided with 8 cores.
GAP8 is an ultra-low-power processor equipped with a built-in hardware Convolutional Neural Network engine. It consumes only milliwatts of power for running typical neural networks.
The typical retail price of a Gapuino GAP8 development board is around USD300.
Myriad X, developed by the company Movidius, is a vision-processing unit with a built-in neural network engine able to deliver up to 1 TOPS (Terflops operations per second)
Myriad X also offers a specialized Deep neural Network engine. The processor can deal easily with memory bottleneck problems.
Myriad X is built on 16 SHAVE cores. SHAVEcores, originally built for game physics engines, have been upgraded as vision accelerators.
One of the interests of the processor is when used in the Neural Compute Stick2 from Intel That stick is usually used for rapid prototyping of edge AI applications. It is plugged into a standard workstation via USB.
The typical retail price of a Neural Compute Stick2 is around USD99
This AI processor from Texas Instrument is part of the Jacinto 7 series which is designed for the automotive industry.
The TDA4VM, based on the Cortex-A72 architecture, is aimed at providing driver assistance functions. For this, it is equipped with a deep learning accelerator chip.
It is able to provide 8 TOPS with a dedicated Matrix Multiplication Accelerator ( typically used by neural networks ) which is a good performance for an IoT processor.
The KL520 is developed by the company Kneron, based in Taiwan. It runs natively Convolutional neural Networks and targets the image recognition market (facial recognition etc)
Its CNN engine can run at 0.3 TOPS which is good enough for most facial recognition applications.
Interestingly enough, the chip architecture can be modified and compression is also used for the optimization of the CNN models.
Google, which provides TensorFlow, the leading AI framework on the market provides AI-based computing cards, the Google TPU, or Tensor Processing Units. These TPUs are available for servers or for edge computing.
A Typical edge TPU can perform at 4 Tera-operations per second.
Tensorflow provides support for TPUs the same way it provides support for GPUs (and ‘traditional’ CPUs). It allows a neural network model to run in a TPU.
TPUs are designed to be connected together to form dedicated AI computing machines.
Here is an illustration of a Cloud using TPU machines:
The typical retail price of a Cloud TPU v2 is around $4.95 / TPU hour
Intel, the famous worldwide microprocessor manufacturer, has clearly decided, during the past recent years, to invest in AI-powered processors.
Mobileye, Agilex FPGA, or Movidius are Intel products that provide AI accelerated processing.
Intel Deep Learning Boost is a new processor technology that increases the speed of execution of AI deep learning for example speech recognition, facial recognition, natural language processing (NLP), etc.
This project of Intel has been stopped.
Intel provides highly-specialized AI processors with hardware routines for training deep neural networks via the Goya and Gaudi processors.
This new Xeon generation is equipped with Deep Learning Boost technology and can provide high performances for neural networks. For example CNNs with the standard neural networks frameworks (Keras+tensorflow / Theano/Caffe)
AMD – same as Intel – has recently invested in AI processors but perhaps with slightly less strength than its rival.
EPYC is a specialized AI processor which can provide up to 5 petaflops of AI power. It can offer 120 cores and it is also used for building powerful AI workstation.
We provided in this article a very brief introductory guide to the recent A.I. cards and processors which can be bought on the market. Most of these A.I. processors allow building modular architecture since they can be scaled in parallel.
Several A.I. workstations are built on these architectures. Their prices range from 10,000$ to 100,000$ or even more.
A.I. is the future and there will be more and more demand for specialized hardware and software which can train and run A.I. applications.
Stay tuned for more news!
Acodez is a renowned web development company India, offering Emerging Technology Services to our clients across the globe. We offer all kinds of web design and web development services to our clients using the latest technologies. We are also a leading digital marketing company providing SEO, SMM, SEM, Inbound marketing services, etc at affordable prices. For further information, please contact us.
Contact us and we'll give you a preliminary free consultation
on the web & mobile strategy that'd suit your needs best.
How Blockchain is Disrupting the World – The Example of the Automotive IndustryPosted on Dec 15, 2020 | Blockchain
Ways Artificial Intelligence Is Improving Software DevelopmentPosted on Dec 10, 2020 | AI and ML
Location-based Marketing: A Guide for 2021 and BeyondPosted on Dec 09, 2020 | Digital Marketing