Driving on the highway, your hands are no longer bound to the steering wheel; instead, you can read a favorite book, watch an entertaining movie with your family, or even take a relaxing afternoon nap… The car is no longer simply a means of transportation, but rather genuinely becomes a mobile leisure space: this is the where the meaning of ‘driverless’ lies. Inspur is helping Baidu to turn the aforementioned scenario into reality within the next 3-5 years.
Does Driverless Car Training Last for a Long, Long Time?
According to US standards, driverless vehicle technology is divided into 5 stages. The current driverless technology research has only reached as far as L4, using artificial intelligence algorithms to achieve completely autonomous driving, mainly relying on high-precision maps, corresponding laser radar, camera, millimeter-wave radar, ultrasonic sensors, GPS and other sensors. Among these, the laser radar scan is equivalent to “eyes”, and is able to scan the surrounding 100-200 meters for objects, the pedestrians, vehicles, traffic signs, distance and other environmental factors in the process of driving to form a real-time road map that is passed to the computing device. The artificial intelligence algorithms serve as the “brain”, providing real-time analysis through the vehicle computing platform, identifying all the data, and making rational judgments to avoid, overtake or whatever course of action suitable for the situation at that time.
In order for driverless cars to have “intelligence”, first of all it is necessary to draw upon the deep learning technology for offline model testing, to make the machine learn through the laser scan and “see” which objects are human, which are animals, which are trees, which are car signals, what traffic signs mean, and so on. However, the current ability of the machine to extract abstract features is far less than that of humans. For example, a 4 or 5 year old child can quickly learn the characteristics of a cat after being exposed to them just a few times, while the Google X lab used more than 16,000 processors, and virtual brains composed of 1 billion nerve nodes to analyze 10 million frames from random untagged Youtube video clips. It took 10 days of operation before the machine could finally distinguish the image of a cat from other frames, and correctly found the cat’s photos from the next input of 20,000 images. The driverless environment is even more complex, and it’s necessary to identify as many as possible things that might be encountered in the process of driving, including a wide variety of people and objects. Such a large learning task of course requires the support of a strong computing power; otherwise the machine may have to learn till the end of the world.
Inspur SR-AI Rack supports 100 billion levels of model training
The offline model training initially used stand-alone multi-card computing devices and began to implement cooperative parallel computing with large-scale GPU clusters as the amount of data increased. However, driverless technology may currently be the most complex application of artificial intelligence, and its model training has already exceeded hundreds of billions of samples, one trillion parameter levels. However, the traditional training tasks are mostly done on a single machine, with only 4-8 cards, and simply cannot satisfy the large storehouse of models and parameters of training performance requirements.
In order to better promote the development of driverless vehicle technology, Inspur and Baidu jointly developed a hyper-scale AI computing module – the SR-AI Rack Scale Server for large-scale data sets and deep neural networks. This product is in accordance with the latest Scorpio 2.5 standards, and is the world’s first AI program using the PCIe Fabric interconnected architecture design. Using the PCI-E switch and I/O BOX modules with GPU and CPU physical decoupling pool, both with flexible configuration, it is able to support 16 GPU hyper scale scalability nodes. At the same time, the SR-AI Rack Scale Server is also the first domestic 100G RDMA GPU cluster. Its supporting RDMA technology (remote data direct access technology) can achieve direct interaction between GPU and memory data, without the need for CPU calculation. It massively reduces server-side data processing delays in network transmission, enabling clusters to reach nanosecond network latency with stand-alone processing capacity of up to 512 TFlops, which more than doubles the performance of conventional GPU servers, and is more than 5-40 times the performance of average AI programs.
Under the new AI computing equipment support, Baidu driverless vehicles have attained an accuracy rate of over 99.9% for traffic light recognition, and an accuracy rate of 95% for identifying pedestrians. In road tests, through GPU and corresponding algorithm support, Baidu driverless cars can accurately identify pedestrians in 0.25 seconds, and with further algorithm optimization in the future, this time will be reduced to 0.07 seconds. Knowing that accidents are inevitable, 0.01 seconds may be the difference between life and death.
In the data center computing, Inspur and Baidu has maintained years of strategic cooperation, do joint develop on artificial intelligence related computing architecture, technology and products aspects, and achieved quite a lot results. Heterogeneous computing server, FPGA acceleration module jointly developed by Inspur and Baidu is widely used in Baidu and other artificial intelligence scene, like Baidu driverless and Baidu Brain.
About Inspur
www.inspursystems.com