nVidia unveiled a platform for unmanned vehicles and the most powerful GPU in the world

The head of Tesla, Elon Musk, appeared on the scene.

On Tuesday, nVidia showed the new most powerful (among systems with one GPU) Titan X video accelerator with 12 gigabytes of memory. The retail cost of new items will be from 999 dollars. Also in the hands of the head of nVidia, Ren-Sun Huang, appeared the Drive PX car computer, which is designed to become the core of autonomous machines.

The new Titan X was shown on March 4, two weeks ago, but that announcement looked crumpled and too empty. At that time, the GDC conference was held, and during the story of representatives of Epic Games about his Unreal Engine game engine and its exactingness for hardware on the scene , Juan suddenly appeared as a special participant . There, the executive director of nVidia talked about the new video card without going into too much detail.

Specific characteristics were not disclosed then; technical publications that could get to know the new card closely were forbidden to distribute any data other than photographs. It was said that there will be 12 gigabytes of video memory, and 8 billion transistors in the chip. Only two weeks later, at the GPU Technology Conference, a full-fledged announcement took place, at which the characteristics and price of the new video accelerator were heard, and immediately after that the Titan X reviews and tests appeared on the network ( 1 , 2 , 3 , 4 , 5 ).



The Titan X is already on sale. The core of the video card is a GM200 chip with 3072 CUDA cores, 192 texture units and 384-bit memory. Above is a simplified device of this chip. On March 4, at the GDC gaming conference, there was an idea that this power is needed for virtual reality helmets, and this Tuesday a lot of time was devoted to machine learning systems and neural networks.

	Titan x	GTX 980	Titan black	R9 290X
GPU	GM200	GM204	Gk110	Hawaii XT
GPU cores	3072	2048	2880	2816
Base frequency	1000 MHz	1126 MHz	889 MHz	1000 MHz
Texture blocks	192	128	240	176
ROP Blocks	96	64	48	64
Memory	12 GB	4 GB	6 GB	4 GB
Memory frequency	7000 MHz	7000 MHz	7000 MHz	5000 MHz
Memory bus	384 bit	256 bit	384 bit	512 bit
Memory bandwidth	336 GB / s	224 GB / s	336 GB / s	320 GB / s
Heat dissipation	250 watts	165 watts	250 watts	290 watts
Peak Performance, Gigaflops	6144	4612	5121	5632
Number of transistors	8.0 billion	5.2 billion	7.1 billion	6.2 billion
Process technology	28 nm	28 nm	28 nm	28 nm
Recommended price	$ 999	549 $	$ 999	359 $
Launch date	2015-03-17	2014-09-18	2014-02-18	2013-10-24

Titan X uses the same Maxwell architecture as the GTX 980, GTX 970, and GTX 960. Roughly speaking, the performance is one and a half GTX 980: 50% more cores, ROPs, texture units, memory bus width, and second-level cache . During the presentation, Juan also mentioned a peak performance of 7 gigaflops, which is different from what is indicated above. This performance is achieved at the highest possible frequency. The predecessors Titan and Titan Black had performance with double calculations three times lower than that of ordinary single. Titan X does not have such a feature, and double-precision calculations are 32 times slower - only 192 gigaflops.

By the way, the memory is not divided into two parts, as happened with the GTX 970. The Titan X has no segmentation, some of which may be slower than others, assures nVidia.

The configuration of the GM200 chip is maximum, there are 24 SMX modules in it, that is, you should not expect another more productive option. The chip is large enough, 8 billion transistors fit in an area of ​​approximately 25 × 25 mm or 625 mm². For comparison: the area considered to be large GK110 (7.1 billion transistors) reached 561 mm². The stock frequencies of the Titan X are slightly (≈13%) lower than that of the GTX 980. The base frequency is 1000 MHz, the maximum reaches 1075 MHz, for 980 these stock values ​​are 1126 and 1216. nVidia claims that overclocking to 1.4 GHz is theoretically possible when cooled by air. The memory frequency is the same - 7 GHz. The most amazing feature of the Titan X is 12 gigabytes of video memory, it's three times the GTX 980 and twice the GTX Titan Black. In a presentation, nVidia addressed 4K gaming issues,



nVidia also talked about how America's number one space company, SpaceX, uses video accelerators in its work. The head of the SpaceX development department, Adam Lichtl, described how the simulation of complex physical models was made possible on video cards. This task could require thousands of conventional processor cores - we are talking about iottabytes (this prefix means 10 ²⁴ ), which are formed when analyzing models of ignition of fuel components.



But without the appropriate software it is impossible to use the power of video cards. nVidia talked about DIGITS, a neural network creation software for researchers. Neural networks can be used to teach object recognition, but their creation often causes difficulties and takes a lot of time. According to nVidia, their product can change everything. DIGITS is available for download at https://developer.nvidia.com/digits . The system has an intuitive interface and supports the version of the Caffe framework, which is processed by the video card.



DIGITS DevBox is the project of the most powerful desktop data thresher. Four Titan X cards are installed in the computer. At the same time, the entire system remains relatively quiet and energy efficient. DevBox comes with preinstalled software products used in neural network research: these are DIGITS, cuDNN 2.0, Caffe, Theano, and Torch. The machine runs under Ubuntu. It is very productive: AlexNet training can take only 13 hours. A system with one video card will require more than 2 days, and on a not-so-weak processor, this task will take more than a month. The cost of DIGITS DevBox is 15 thousand dollars - this is not a gaming computer for Crysis, but scientific equipment.







Ren-Sun Huang also discussed the future architecture of Pascal. It will be based on the 16-nm FinFET + process technology - nVidia passes 20 nm. Pascal's performance per watt will be more than two times higher than that of Maxwell. Another major improvement is the use of a more productive memory called High Bandwidth Memory. nVidia claims that up to 32 GB of memory per GPU will be available, and throughput will triple. Thus, up to a terabyte per second bandwidth is theoretically possible.



Pascal-based cards will use NVLink for the first time, a high-speed bus between the GPU and GPU, or between GPUs. NVLink speed is significantly higher than PCI Express. The memory of the new architecture will use 3D technology, that is, the chips will also have a vertical orientation. As a result of all these innovations, Pascal can process some processes 10 times faster. Particular attention is paid to optimizing the tasks of machine vision, image recognition, construction and functioning of neural networks. Juan showed Drive PX and named its price. This is the car’s autopilot computer, which was first talked about.



back at CES 2015 in January of this year. Two Tegra X1 mobile chips are installed on the board with a total performance of 2.3 teraflops, which are able to process in real time the video stream from 12 different HD cameras using 630 million neural network connections. This computer is designed for self-learning ADAS systems and future unmanned vehicles.

Drive PX comes with the DIGITS software mentioned above, as well as video capture and processing libraries. The product is intended for both real car manufacturers and research projects. Autonomous cars are the future, but the nVidia computer can bring it closer. Drive PX sales will begin in May, the cost of the platform is 10 thousand dollars.



Shortly before the announcement of Drive PX, Ilona Mask was invited to the scene. He discussed machine vision and autonomous cars. Juan stabbed Mask: the head of Tesla is afraid of artificial intelligence , but at the same time believes that unmanned vehicles are much safer than human drivers. Musk explained his position: danger is only potential, and an autonomous machine has a narrow form of AI. Musk is confident that in the future we will look at the need to drive our cars in the same way as today we perceive the uselessness of the operator in the elevator. Unfortunately, no words were heard about the collaboration between nVidia and Tesla.



Based on materials from ExtremeTech ( 1 , 2 ), the nVidia blog ( 1 , 2 , 3 , 4 ), Re / code , PC Perspective , Techgage , HotHardware ( 1 , 2 , 3 , 4 ), ITworld , CNET , AnandTech, and The Register .