• This email address is being protected from spambots. You need JavaScript enabled to view it.

ADxxxx - L4/L5 Autonomous Drive Platforms

L5 autonomous driving = the vehicle has the same mobility as a human driver - unlimited!
The vehicle must adjust to all driving scenarios. All road, weather and traffic conditions.

A daunting taks requiring the ability to perceive a large amount of sense information, fusing this to make decisions based on past experience and imagination - all in milliseconds!

A combination of AI and advanced signal processing (mathematical) algorithms in multiple stages processed efficiently at high speed is required.

Why Use Both AI and Advanced Signal Processing?

L4/L5 Sensor Fusion

L4/L5 Sensor Fusion

At Level 4 and Level 5 it is no longer feasible to rely on the fusion of individiual sensor group. In order to get a proper understanding of the environment there will be a need to combine fusion between sensor groups as well as the type of sensors.

Consider the combination of cameras and lidar for example. The cameras would typically be using AI algorithms to extract relevant information. If the camera information is fused with lidar data a more correct 3D image of the relevant space can be formed.

But the lidar data can also be fused with the data from the radar. In this case advanced mathematical calculations / signal processing will be applied to extract the interesting information.

For L4/L5 it is not enough to handle the fused data separately, the AI and DSP fused data must also be combined together with data from the GNSS unit and the IMU. At this stage enough information is available to create a relevant input to the planning/decision stage.

 

AD Planning Stage

AD Planning Stage

The planning stage is usually broken down hierarchially so that decisions that are needed to be taken can be done so based on relevant information. To reach the decision point you need to apply a combination of mathematical models and AI processing.

In a typical scenario the planning / decision process will pass through three stages, where the first stage would be the mission planning stage.

In mission planning the aim is to find the shortest path between two destinations. The output from this stage would be a high definition map with a mission path completed with localization information. This output will feed the behavioural planning stage, which is next.

Behavioural planning looks at the rules of the road, what objects are on the way - both static and dynamic - and the aim is to find a safe and efficient path to the target in the form of the driving manoeuvre to execute and what constraints to follow.

Local planning uses the path and the velocity profile to look at trajectory and consider elements like potential collisions and acceleration constraints to provide a trajectory, path and velocity output to the vehicle control stage.

The optimal output is provided by using a combination of mathematical models and AI processing.

ADxxxx Family

This image for Image Layouts addon
The ADxxxx family consists of a series of computational platforms configured as co-processors to the main processor. These allow for the design of an extremely compact and low-power single chip L4/L5 solution.
ADxxxx Family Block Diagram

ADxxxx Family Block Diagram

There are 3 main components to the ADxxxx family:

  • An advanced signal processing unit (DSP)
  • An AI unit consisting of a series of identical cores
  • A tightly coupled memory (TCM).

Advanced Signal Processing Unit

The signal processing part of the ADxxxx contains a number of ALUs - starting with 512 in the AD0514. This is a normal signal processing unit, with the main difference that the number of available ALUs are significantly higher than what can be found in other DSPs of today. The other main difference is that the DSP can access all of the available TCM.

AI Unit

The AI unit consists of a series of cores. Each core is built up of 16k MACs. The AD0514 contain 8 cores, which provides a total of 128k MACs to the system. The AI unit can access all of the TCM as well, and thus the data can quickly and efficiently be shared between DSP and AI unit.

TCM

The size of the TCM is decided by the designer. Optimal size can be derived using the development tools, but even if the size of the TCM should be too small for optimal performance the unit will still work, only with a slowre throughput.

Performance Examples AD1028

Clock frequency 2GHz

Performance Examples 

514 Tflops

AD0514

  • 512 ALUs in a single core providing 2 Tflops @ 2GHz
  • 128k MACs divided into 8 cores, each with 16k MACs providing 512 Tflops @ 2GHz
  • User configurable memory size
  • IEEE754 floating point with user selectable accuracy (exponent and mantissa)
  • High-level development environment

2056 Tflops / 2 Petaflops

AD2056

  • 2048 ALUs in a single core providing 8 Tflops @ 2GHz
  • 512k MACs divided into 32 cores, each with 16k MACs providing 2048 Tflops @ 2GHz
  • User configurable memory size
  • IEEE754 floating point with user selectable accuracy (exponent and mantissa)
  • High-level development environment