A bit of history..
It’s no wonder that many years ago, about 6 (in computer terms, that is) some companies started to provide specialized hardware & Software solutions to improve the performance of AI and Machine Learning algorithms, like nvidia with its CUDA platform. This is has been really important in the AI/ML industry as this graph shows:
Basically, an improvement of 33 times the speed of using a normal pc..
But if this graphic was not enough to motivate you to learn more (and get to the end of this article) – see this other one:
This is a graph made on 2016 showcasing the evolution regarding AI processing power since 2012, which the 1X at the bottom is based on an already accelerated GPU for AI processing… which was set as a landmark or baseline on 2012 with Alex Krizhevsky’s study regarding usage of a Deep convolutional neural network that learned automatically to recognice images from 1 million examples. With only two days of training using two NVIDIA GTX 580 GPUs. The study name was “ImageNet Classification with Deep Convolutional Neural Networks”
It’s a BANG! – A big one, which many are calling the new industrial revolution – AI. There, many companies listened, adopting this technology: Baidu, Google, Facebook, Microsoft adopted this for pattern recognition and soon for more..
Between 2011 and 2012, a lot of things happened on AI: Google Brain project achieved amazing results – being able to recognize cats and people by watching movies (though using 2,000 CPU at Google’s giant data center) – then this result was achieved by just 12 NVIDIA GPUs This feat was performed by Bryan Catanzaro from NVIDIA along (my teacher!) Andy NG’s team at Stanford (Yay! I did your course so I can call you teacher :D)
Later on 2012, Alex Krizhevsky from the University of Toronto won the 2012 ImageNet computer image recognition competition, by a HUGE margin, beating image recognition experts. He did NOT write computer vision code. Instead, using Deep Learning, his computer learned to recognice images by itself, they named their neural network AlexNet and trained it with a million example images. This AI bested the best human-coded software.
The AI race was on…
Later on, by 2015, Microsoft and Google beat the best human score in the ImageNet challenge. This means that a DNN (Deep Neural Network) was developed that bested human-level accuracy.
2012 – Deep Learning beats human coded software.
2015 – Deep Learning achieves beats human level accuracy. Basically acquiring “superhuman” levels of perception.
To have an idea, the following graphic shows the acquired accuracy of both Computer Vision and Deep Learning algorithms/models:
Related to this, I wanted to highlight the milestone achieved by Microsoft’s research team on 2016 but before this, let me mention what Microsoft’s chief scientist of speech, Xuedong Huang said on December 2015: “In the next four to five years, computers will be as good as humans” at recognizing the words that come of your mouth.
Well, on October 2016, Microsoft announced a system that can transcribe the contents of a phone call with the same or fewer errors than actual human professionals trained in transcription… Again human perception has been beaten..
The Microsoft research speech recognition team
These advancements are made possible by the improvement in Deep Learning mainly which is acquired by massive calculation power like 2.000 servers of Google Brain or, as of now, just a few NVIDIA GPUs… this delivers results and results drive the industry and make it trust a technology and, more importantly, bet on it. This is what is has been happening along this years…
Our current AI/ML/DL “Boosters”:
They are essential tools to boost AI (ML, Deep Learning, etc..) and are supported by a day by day increasing number of tools and libraries.. (Caffe, Theano, Torch7, TensorFlow, Keras, MATLAB, etc..) and many companies use them (Microsoft, Google, Baidu, Amazon, Flickr, IBM, Facebook, Netflix, Pinterest, Adobe,… )
An example of this is the Titan Z with 5,760 CUDA cores, 12GB memory and 8 Teraflops
Comparatively, “Google Brain” has 1 billion connections spread over 16,000 cores. This is achievable with $12K with three computers with Titan Z consuming “just” 2,000 KW of power, Ditto.. – oh and if this sounds amazing, this is data from 2014… yeah, I was just teasing you 😉
It gets better…
As of today, we have some solutions already on the consumer market, which you might have in your home computer, like the NVIDIA Pascal based graphic cards:
Nvidia 1080 with 10 Gbps, 2560 NVIDIA CUDA Cores and 8GB GDDR5X memory
NVIDIA Titan Xp with 11 Gbps with 3584 NVIDIA CUDA cores and 12 GB GDDR5X memory
Here is a picture of the beautifully crafted NVIDIA 1080, launched by the end of June 2016:
And it’s my current graphic card, from when I decided to focus on Machine Learning and Data Science, by the end of 2016 😉 – I am getting ready for you baby! (currently learning Python)
Also, similarly, we have the Quadro family, focused on professional graphic workstations, for professional use. Being their flagship the Quadro P6000 with 3840 CUDA cores 12 Teraflops and 24GB GDDR5X.
And this just got better and better…
I could not help myself reminding myself of this scene from Iron Sky
Recently announced this past 10th of October 2017 we have the Pegasus nvidia drive PX, the autonomous supercomputer for fully autonomous driving, with a passively cooled 10 watts mobile CPU () with four high performance AI processors. Altogether they are able to deliver 320 Trillion Operations per Second (TOPS)
Pegasus! – I personally love the name (I think Mr. Jensen Huang must like the “Zodiac Cavalliers” very much! – as a good geek should ^.^)
I believe these AI processors are two the newest Xavier system-on-a-chip processors coupled with an embedded GPU based on the NVIDIA Volta architecture. The other two seem to be two next generation discrete GPU with hardware explicitly created for accelerated Deep Learning and computer vision algorithms. All in the size of a license plate.. not bad!
Here is a pic of the enormous “Pegasus” powerhorse:
This is huge – again yeah. Think that this is basically putting 100 high-end servers in the size of a license plate.. Servers on current Hardware, that is..
And this is powered by…
Did I say volta?
This is nvidia’s GPU Architecture which is meant to bring industrialization to AI, and has a wide range of their products supporting this platform. NVIDIA Volta is meant for healthcare, financial, big data & gaming..
This hardware architecture consist of 640 Tensor cores which deliver over 100 Teraflops per second, 5x the previous generation of nvidia’s architecture (Pascal).
DGX systems – AI Supercomputers “a la carte” Based on the just mentioned Volta architecture, having 4x TESLA V100 or the Rack based supercomputer DGX-1 with up to 8 TESLA V100, having each an intel Xeon for each 4 V100. Oh, and all the other hardware boosted to support these massive digital brainpower..
Following some comparative picture to put things in the proper perspective…
Here, in the hands of Jensen Huang, who is Nvidia co-founder and CEO, is a Volta V100, if you were wondering:
Smaller than the 100x servers it can beat, right?
V100 family, along Volta Architecture, were presented just recently this year at Computex, end of May.
Oh, and the market responded extremely well…
They are also empowering IOT solutions for embedded systems, targeting small devices like drones, robots, etc.. to perform video analytics and autonomous AI, which is started becoming a trend now in consumer products..
The family of these products is called NVIDIA Jetson, with its TX2 being their flagship, having 256 CUDA cores and 8GB 128 bit LPDDR4 memory along two CPU (HMP Dual Denver + Quad ARM)
As you can see the race is on, and continues to accelerate and who knows where it will bring us to..
Hope you enjoyed this post, if you liked it, please subscribe 🙂
So, what do you think?
Please respond directly on my blog so I do not have to work on recopilating the information from different sources..
One thought on “Give me power, Pegasus! – or the state of Hardware in AI”