CGOCmall - Tesla's New Patent Kills Nvidia, The First Autopilot Chip To Abandon Cpu And Gpu

Welcome to CGOCmall ! | Register

Home > Industry information > Tesla's New Patent Kills Nvidia, The First Autopilot Chip To Abandon Cpu And Gpu

Tesla's New Patent Kills Nvidia, The First Autopilot Chip To Abandon Cpu And Gpu

Auth:CGOCMALL Date:2019/1/28 Source:CGOCMALL Visit:412 Related Key Words: Tesla Nvidia chip CGOCMALL
In the field of autonomous driving, GPU's powerful computing power is an option for autopilot chips, and it has also created NVIDIA's position in this field with the most powerful GPU. NVIDIA Xavier may also become the world's first automatic ISO26262 standard. Driving the chip.

However, the artificial intelligence chip that Tesla is developing, also known as the neural network chip, can also bring the complete car to the car without using the CPU and GPU. Autopilot capability, this chip will be released with the Autopilot 3.0 hardware version.

In Tesla's latest published patent application, explains why Tesla gave up CPU and GPU to provide power for Tesla's machine learning system:

The processing of machine learning and artificial intelligence usually requires mathematical operations on large amounts of data. Usually, multiple convolutional layers and pooling layers need to be solved. Machine learning and artificial intelligence techniques usually use nonlinear functions such as matrix operations and activation functions. Machine learning Applications include autonomous driving and assisted driving.

In some scenarios, a computer processor (CPU) is used to perform machine learning training and reasoning. Traditional computer processors can perform a single mathematical operation very quickly, but usually only a limited amount of data can be processed simultaneously. As an alternative, a graphics processing unit (GPU) can be used and the same mathematical operations can be performed on larger data sets in parallel.

By using multiple processor cores, the GPU can perform multiple tasks in parallel, typically performing large graphics processing tasks faster than traditional computer processors. However, GPUs and traditional computer processors were not originally designed for machine learning or artificial intelligence operations.

Machine learning and artificial intelligence operations typically rely on repeatedly applying a specific set of machine learning processor operations on very large data sets. Therefore, there is a need for a microprocessor system that supports parallel execution of machine learning and artificial intelligence specific processing operations on large data sets without the need for multiple processing cores for each parallel operation.

The patents nominated by Tesla's patents include Peter Bannon, former Apple chip architect and current Tesla chip project leader; Emil Taples, who has long served as AMD chip architect; Debjit Das Sarma, former AMD chief CPU architect.

Tesla's series of patents describe the microprocessors designed to solve the problem of abandoning GPUs. The following is an introduction to Tesla's new artificial intelligence chip patent:

01

Accelerated digital engine

Tesla describes the invention in a patent application:

Various embodiments of the present invention are related to an accelerated math engine: in some embodiments, the accelerated math engine is applied to image processing to process by using a two-dimensional matrix comprising sub-circuits including ALUs, output registers, and shadow registers. Accelerate the convolution of the image. This architecture supports a clocked two-dimensional architecture in which image data and weights are multiplied in a synchronized manner, allowing a large number of mathematical operations to be performed in parallel.

The following are some drawings and schematics when applying for a patent:

The following are some drawings and schematics when applying for a patent | CGOCMALL
02

Variable delay memory access computing array microprocessor system

Tesla describes the invention in a patent application:

The microprocessor system includes a computing array and a hardware arbiter: the computing array includes a plurality of computing units. Each of the plurality of computing units operates on a respective value addressed from the memory. The hardware arbitration program is configured to control at least one memory request to issue one or more corresponding values from the memory of the computing unit. The hardware arbiter is also configured to schedule a control signal to be issued based on the issuance of the memory request.

The following are some drawings and schematics when applying for a patent:
Tesla's new patent kills Nvidia, the first autopilot chip to abandon CPU and GPU | CGOCMALL
03

Computational array microprocessor system for non-continuous data formats

Tesla describes the invention in a patent application:

The microprocessor system includes a computing array and a hardware data formatter: the computing array includes a plurality of computing units, each computing unit operating on a corresponding value addressed from memory. The values operated by the computing unit are provided to the computing array as a set of values to be processed in parallel. The hardware data formatter is configured to collect a set of values, where the set of values includes a first subset of values placed consecutively in memory and a second subset of values placed consecutively in memory. There is no need to contiguously locate the first subset of values in memory from the second subset of values.

The following are some drawings and schematics when applying for a patent:
Tesla's new patent kills Nvidia, the first autopilot chip to abandon CPU and GPU

04

Vertor calculation unit

Tesla describes the invention in a patent application:

Tesla's new patent kills Nvidia, the first autopilot chip to abandon CPU and GPU | CGOCMALL

The microprocessor system includes a computing array and a vector computing unit: the computing array includes a plurality of computing units. The vector calculation unit is in communication with the computing array and includes a plurality of processing elements. The processing element is configured to receive the output data elements of the computed array and process the received output data elements in parallel.

The following are some drawings and schematics when applying for a patent:

The design of Tesla's autopilot chip provides a new direction for solving the problem of autonomous driving, which may be the most worthwhile study of this chip. The reason for adopting a microprocessor-based design approach is perhaps because the design is more in line with Tesla's automated driving route hardware.

Industry information

Product Index :

客户服务
live chat
客服系统
live chat