Monday, January 4th 2016

NVIDIA Announces Drive PX 2 Mobile Supercomputer

NVIDIA announced the Drive PX 2, the first in-car AI deep-learning device. This lunchbox sized "mobile supercomputer" embeds up to twelve CPU cores, a "Pascal" GPU built on the 16 nm FinFET process, 6 TFLOP/s of raw compute power, and 24 deep-learning TOps of compute power usable for deep-learning applications; the chips are liquid-cooled, draw 250W in all, and give the car a very powerful deep-learning device for self-driving cars. The device itself will be offered to car manufacturers to redesign and co-develop self-driving cars with.
The press-release follows.

Accelerating the race to autonomous cars, NVIDIA (NASDAQ: NVDA) today launched NVIDIA DRIVE™ PX 2 -- the world's most powerful engine for in-vehicle artificial intelligence.

NVIDIA DRIVE PX 2 allows the automotive industry to use artificial intelligence to tackle the complexities inherent in autonomous driving. It utilizes deep learning on NVIDIA's most advanced GPUs for 360-degree situational awareness around the car, to determine precisely where the car is and to compute a safe, comfortable trajectory.

"Drivers deal with an infinitely complex world," said Jen-Hsun Huang, co-founder and CEO, NVIDIA. "Modern artificial intelligence and GPU breakthroughs enable us to finally tackle the daunting challenges of self-driving cars.
"NVIDIA's GPU is central to advances in deep learning and supercomputing. We are leveraging these to create the brain of future autonomous vehicles that will be continuously alert, and eventually achieve superhuman levels of situational awareness. Autonomous cars will bring increased safety, new convenient mobility services and even beautiful urban designs -- providing a powerful force for a better future."

24 Trillion Deep Learning Operations per Second
Created to address the needs of NVIDIA's automotive partners for an open development platform, DRIVE PX 2 provides unprecedented amounts of processing power for deep learning, equivalent to that of 150 MacBook Pros.

Its two next-generation Tegra® processors plus two next-generation discrete GPUs, based on the Pascal™ architecture, deliver up to 24 trillion deep learning operations per second, which are specialized instructions that accelerate the math used in deep learning network inference. That's over 10 times more computational horsepower than the previous-generation product.

DRIVE PX 2's deep learning capabilities enable it to quickly learn how to address the challenges of everyday driving, such as unexpected road debris, erratic drivers and construction zones. Deep learning also addresses numerous problem areas where traditional computer vision techniques are insufficient -- such as poor weather conditions like rain, snow and fog, and difficult lighting conditions like sunrise, sunset and extreme darkness.

For general purpose floating point operations, DRIVE PX 2's multi-precision GPU architecture is capable of up to 8 trillion operations per second. That's over four times more than the previous-generation product. This enables partners to address the full breadth of autonomous driving algorithms, including sensor fusion, localization and path planning. It also provides high-precision compute when needed for layers of deep learning networks.

Deep Learning in Self-Driving Cars
Self-driving cars use a broad spectrum of sensors to understand their surroundings. DRIVE PX 2 can process the inputs of 12 video cameras, plus lidar, radar and ultrasonic sensors. It fuses them to accurately detect objects, identify them, determine where the car is relative to the world around it, and then calculate its optimal path for safe travel.

This complex work is facilitated by NVIDIA DriveWorks™, a suite of software tools, libraries and modules that accelerates development and testing of autonomous vehicles. DriveWorks enables sensor calibration, acquisition of surround data, synchronization, recording and then processing streams of sensor data through a complex pipeline of algorithms running on all of the DRIVE PX 2's specialized and general-purpose processors. Software modules are included for every aspect of the autonomous driving pipeline, from object detection, classification and segmentation to map localization and path planning.

End-to-End Solution for Deep Learning
NVIDIA delivers an end-to-end solution -- consisting of NVIDIA DIGITS™ and DRIVE PX 2 -- for both training a deep neural network, as well as deploying the output of that network in a car.

DIGITS is a tool for developing, training and visualizing deep neural networks that can run on any NVIDIA GPU-based system -- from PCs and supercomputers to Amazon Web Services and the recently announced Facebook Big Sur Open Rack-compatible hardware. The trained neural net model runs on NVIDIA DRIVE PX 2 within the car.

Strong Market Adoption
Since NVIDIA delivered the first-generation DRIVE PX last summer, more than 50 automakers, tier 1 suppliers, developers and research institutions have adopted NVIDIA's AI platform for autonomous driving development. They are praising its performance, capabilities and ease of development.

"Using NVIDIA's DIGITS deep learning platform, in less than four hours we achieved over 96 percent accuracy using Ruhr University Bochum's traffic sign database. While others invested years of development to achieve similar levels of perception with classical computer vision algorithms, we have been able to do it at the speed of light."

-- Matthias Rudolph, director of Architecture Driver Assistance Systems at Audi
"BMW is exploring the use of deep learning for a wide range of automotive use cases, from autonomous driving to quality inspection in manufacturing. The ability to rapidly train deep neural networks on vast amounts of data is critical. Using an NVIDIA GPU cluster equipped with NVIDIA DIGITS, we are achieving excellent results."

-- Uwe Higgen, head of BMW Group Technology Office USA
"Due to deep learning, we brought the vehicle's environment perception a significant step closer to human performance and exceed the performance of classic computer vision."

-- Ralf G. Herrtwich, director of Vehicle Automation at Daimler
"Deep learning on NVIDIA DIGITS has allowed for a 30X enhancement in training pedestrian detection algorithms, which are being further tested and developed as we move them onto NVIDIA DRIVE PX."

-- Dragos Maciuca, technical director of Ford Research and Innovation Center
The DRIVE PX 2 development engine will be generally available in the fourth quarter of 2016. Availability to early access development partners will be in the second quarter.
Add your own comment

31 Comments on NVIDIA Announces Drive PX 2 Mobile Supercomputer

#1
zithe
Now my car will tell me to stop picking my nose.
Posted on Reply
#3
Mistral
"World's First AI Supercomputer," this?

Talk about PR department on steroids.
Posted on Reply
#4
xvi
I hope they're planning on using a separate cooling loop. If they're thinking about tying that in to a car's cooling system, they're going to be dealing with ~90c water temps as-is.

Edit: Are any of you WCG or F@H members thinking what I'm thinking?
Posted on Reply
#5
HumanSmoke
zithe said:
Now my car will tell me to stop picking my nose.
For the Audi's, BMW's, and Daimlers the DNN calculates the optimum nose picking experience
For the Ford, the DNN calculates how you can juggle nose picking, picking your fantasy football team, and getting to the nearest McDonalds.
Quick screengrab of the Pascal GPUs on the reverse side of the PX2 unit


and the obligatory numbers slide...


Volvo announced as the launch partner for the system.
Posted on Reply
#6
FordGT90Concept
"I go fast!1!11!1!"
xvi said:
Edit: Are any of you WCG or F@H members thinking what I'm thinking?
The slide shows 8 TFlOp versus 7 TFlOp of Titan X so, no. This wouldn't be bad for those workloads but most of its hardware would go to waste too.
Posted on Reply
#7
HumanSmoke
FordGT90Concept said:
The slide shows 8 TFlOp versus 7 TFlOp of Titan X so, no. This wouldn't be bad for those workloads but most of its hardware would go to waste too.
I'm guessing that as an automotive part the replacement pricing will be out of all proportion with its actual cost of manufacture ;). GDDR5 chips and dual GPUs in less than a 250W power envelope tends to indicate a mid/lower tier GPU, so a repeat of Kepler and Maxwell? GP104/106 first for consumers, GP100 initially for Tesla.
Posted on Reply
#8
Haytch
Still a far cry from Knight Industries Two Thousand.
Posted on Reply
#9
Tsukiyomi91
as powerful as 150 Macbook Pro... Apple has got to step up their self-driving cars project if they want to keep up with Nvidia as they have not produce a hardware the size of a lunchbox to house a specialized AI that can learn, think & reason yet despite they got more bank than a company that sells pixel pushers. Same goes to Google as they too do not have a proper hardware to fight what Nvidia has unveiled at the recent keynote, named the DRIVE PX 2.
Posted on Reply
#10
Xzibit
Tsukiyomi91 said:
as powerful as 150 Macbook Pro... Apple has got to step up their self-driving cars project if they want to keep up with Nvidia as they have not produce a hardware the size of a lunchbox to house a specialized AI that can learn, think & reason yet despite they got more bank than a company that sells pixel pushers. Same goes to Google as they too do not have a proper hardware to fight what Nvidia has unveiled at the recent keynote, named the DRIVE PX 2.
The actually learning happens off-site not in the cars computer. Atleast according to what they told Ryan over at PCPerspective.
Posted on Reply
#11
Tsukiyomi91
I think putting a special box with large capacity HDDs to let the AI "sit" in the car to drive, index addresses, locations & whatnot would be a more realistic option IMO... much better than over a wireless network where constant update can prove to be a hassle for some.
Posted on Reply
#12
Xzibit
Tsukiyomi91 said:
I think putting a special box with large capacity HDDs to let the AI "sit" in the car to drive, index addresses, locations & whatnot would be a more realistic option IMO... much better than over a wireless network where constant update can prove to be a hassle for some.
That's what all driver-less cars have been doing so far. I haven't seen the Press-Conference but if its anything like the past 2yrs Nvidia is pushing for workload to drive more sensors on the car.
Posted on Reply
#14
Recus
Thanks for driving evolution, Nvidia.
Posted on Reply
#15
HumanSmoke
Tsukiyomi91 said:
I think putting a special box with large capacity HDDs to let the AI "sit" in the car to drive, index addresses, locations & whatnot would be a more realistic option IMO... much better than over a wireless network where constant update can prove to be a hassle for some.
The idea is not to evolve the neural net in isolation but for all vehicles using the system to upload to the server so that every system connected to it gains from the combined data input/experience of all inputs. Every input from every vehicle adds to the knowledge base - the neural net develops faster, and updates can be applied to vehicles that have not encountered the situations personally. It also means that using this method, new vehicles off the showroom floor don't start out relatively "dumb" since they have the cumulative deep learning experience of thousands or millions of other vehicles.
Posted on Reply
#16
okidna
Xzibit said:
The actually learning happens off-site not in the cars computer. Atleast according to what they told Ryan over at PCPerspective.
Yes, the DNN training ("learning") process will be off-site (and also on cloud using DIGITS) because it will take hours or even days to complete the training process.

The Drive PX2 will use the trained weight data to do online/direct image detection and classification for inference purposes (identifying road signs, avoiding obstacles, avoiding pedestrians, avoiding other vehicles, etc.), bear in mind that there will be hundreds or thousands of images to be processed each second so don't be surprised about the high GPU capability and "2,800 images/sec." claim. Yes, hardware wise it might go to waste, but in an autonomous driving it's always better be safe than sorry.

Oh and also it can be used to collect extra and new images and sensor data for further training purpose.

HumanSmoke said:
The idea is not to evolve the neural net in isolation but for all vehicles using the system to upload to the server so that every system connected to it gains from the combined data input/experience of all inputs. Every input from every vehicle adds to the knowledge base - the neural net develops faster, and updates can be applied to vehicles that have not encountered the situations personally. It also means that using this method, new vehicles off the showroom floor don't start out relatively "dumb" since they have the cumulative deep learning experience of thousands or millions of other vehicles.
That's more like it :)

The key is cumulative but localized training.
Cumulative means that the training process will get as much data as it want, so the model can minimized overfitting, and localized means that all the data you get will be relevant and not redundant.
Posted on Reply
#17
cyneater
Hopefully it can address all its ram :P ?
Posted on Reply
#18
the54thvoid
As @HumanSmoke says, the system pools knowledge and all units gather data, process and resolve. It's how Google's cars are all running. Novel stimuli get uploaded so all units can learn collectively.

Resistance is Futile.
Posted on Reply
#19
truth teller
mandatory liquid cooling? why not make the enclosure a passive cooler with some fins? are vehicle manufacturers ok with having to add another coolant loop (i mean this _has_ got to be on its own separate loop, pipes+reservoir+pump+radiator, or are they passing engine coolant [can reach 100ºc easily] throught this)

maybe if they can train the system to park in the garage for me it might have some usefullness. leave car in drive way, walk away, car opens garage door waits for it to rise, enter garage, commands garage door to lower and powers down


Tsukiyomi91 said:
there's always room for AI to grow.
oh, and also, this isnt ai, its machine learning, although they might seem similar to most they are very different
Posted on Reply
#20
Lionheart
So can my car run Crysis? :pimp:

I'll leave.....
Posted on Reply
#21
xvi
Tsukiyomi91 said:
as powerful as 150 Macbook Pro
..which is similar to saying the car is as powerful as 150,000 gerbils (about 200 horsepower, by the way).
Posted on Reply
#22
HumanSmoke
xvi said:
..which is similar to saying the car is as powerful as 150,000 gerbils (about 200 horsepower, by the way).
Don't let the Gerbil Liberation Army hear you talk like that

Posted on Reply
#23
ur6beersaway
So when I overclock this thing, will I receive a speeding ticket? :)
Posted on Reply
#24
xvi
ur6beersaway said:
So when I overclock this thing, will I receive a speeding ticket? :)
Only if they can catch you. :P
Posted on Reply
#25
Vayra86
ur6beersaway said:
So when I overclock this thing, will I receive a speeding ticket? :)
As long as you don't touch the throttle... all is fine :P
Posted on Reply
Add your own comment