So far, from Nvidia, Mobileye and NXP chip supplier information appear to have their own concept of automatic driving vehicle platforms (and how they are going to achieve). Given that everyone use their existing, and they think they can beat his opponent to preempt the market position, it's understandable.

But it is worth noting that the automobile factory and a line of automobile components supplier, their challenge is the same: the car more and more the number of the electronic control unit (ECU), automatic vehicle with various sensors, the sensor data need to be collected processing, analysis and integration, and security issues ─ ─ networking cover door of car.

Those challenges and advanced visual processing technology, deep learning, mapping and so on function, also can affect the new system architecture of processor performance requirements.

This is Google self-driving vehicles inside look like?

So, there is a problem worth $64 million ─ ─ today's automobile manufacturers and a line of automobile components supplier, already know 2020 automatic driving system architecture?

French Kalray new IC design company executive EricBaissus recently an interview with EE Times editor, answer to these questions is: they don't know, or don't know; And this is why the start-up company thinks that its 288 core of large scale Parallel VLIW Processor Array (Massively Parallel Processor Array, MPPA), has arrived at a good time to enter the market.

Kalray was originally for the French atomic energy commission (CEA), the limits of the computing technology needed to develop nuclear bullet simulation; And the company is now a lock key embedded market (such as space), and cloud computing.

Baissus think, self-driving vehicles also belongs to the key part of the embedded market, because such vehicles need to absorb a large number of data from the car, the car parts, rapid processing, then to make quick decision; He said, so the car industry: "the need to deal with multiple domain function integration (multi - domain function integration) and in high level processing tasks of a new generation of processors."

Of course, many of the so-called "core (manycore) revolution" has begun; But Baissus said: "no one has been successfully designed a large scale parallel, has more than 100 core" super single chip computer ";" Bostan Kalray the latest generation of 288 core processor, integrated the 16 17 core processor cluster, have their own equipped with 2 MB of Shared memory (SMEM), the data transmission speed 80 gb per second, and there are 16 core system.

In addition, Bostan also is able to adapt to the critical moment of network single chip, support high-speed Ethernet network interface (x1 GbE 8 ~ 10 GbE). The chip and equipped with high speed encryption and decryption, and the simple links with the GPU/FPGA accelerator function.

So the MPPA architecture can provide the speed of the DSP type, save electricity, temporal predictable (timing predictability), multi-domain support (for example, different processor cluster can perform the car used by different parts of different embedded systems), and scalable parallel computing (internal processor can be together in response to a system of complexity).

This for self-driving vehicles "super computer" single chip, with the Nvidia Drive PX platforms wouldn't it be like? Nvidia will Drive the PX is referred to as "the world's most advanced automatic driving vehicle platforms," which the platform is a combination of deep learning, sensors, panorama video etc. Function.

And Baissis explanation, the differences between the two are two: first, Kalray solution is "certification (certifiable)" : "what I mean is that we can prove that determinism (determinism), and can guarantee the time sequence; in high-performance computing, 1 second delay can be accepted, but in key embedded markets ─ ─ such as aerospace and automobile ─ ─ just 10 milliseconds (millisecond delay is likely to be fatal."

Second, he said engineers need to use the Nvidia chip understand CUDA, but: "our chips can use standard tools and Linux implementation of standard C/C + + program code;" Automobile manufacturers have a lot of the old code of C language to write and algorithm, even if automakers to adopt a new automatic driving vehicle platforms, old code is still very important.

Not only Nvida expected future cars will require more processing performance, another chip industry Mobileye also in recent advance "published" EyeQ5 processor, and promised in 2018 to provide engineering sample chip.

EyeQ5 adopts advanced under 10 nanometers or FinFET process to carry on the design, will be equipped with eight multithreaded processor core, as well as a new generation 18 Mobileye vision processor core; The company said the EyeQ5 can perform 12 Tera per second, at the same time to control power consumption under 5 w.

While everyone including Baissus not underestimated Mobileye; Different from Nvidia Drive PX by many industry watchers as a test platform, automatic driving vehicle Mobileye follow the commercial market in demand higher processing performance, also requires a lower level of power consumption trend; Visual processing algorithm through the use of proven, EyeQ5 data fusion ─ ─ combined with 20 external processor such as camera, radar, lidar ─ ─ cover in a single chip.

But EyeQ5 can run within the vehicle ECU automatically? A Mobileye spokesman for EE Times explanation, EyeQ5 not only support the data fusion, also can perform the decision, but the decision into action is performed in other ways ─ ─ namely car makers have chosen low-order ECU.

And Kalray many of its core processor role positioning, and Mobileye and Nvidia is slightly different. Baissus said that in terms of sensor and machine learning algorithms, there has been a lot of it is very necessary for automatic vehicle progress: "but there is no real as in the field of the processor is;" This is he saw an opportunity.

Baissus believes that a new generation of automatic processor to execute a vehicle beyond the function of the data fusion: "they must be more open platform;" And he is expected to provide an automatic vehicle open processing center ─ ─ could be called "super ECU". The super ECU can on a single chip to provide interdisciplinary integration function, to include sensors, learning, and the key elements such as security, network and cost to bring better results.

Did not disclose the manufacturer name, Baissus said a leader and a line of major car companies auto parts suppliers, are adopting Kalray platform to build the first prototype self-driving car, but he also acknowledged that current automatic vehicle structure is not mature enough; However, through the cooperation with several major vendors, he was eager to know more about the demand of the depot, to help the company define the next generation of self-driving vehicles solution. While Baissus does not rule out the MPPA architecture licensed to other automotive chip manufacturers.