Tesla Develops Own Self-Driving AI Chip - Removes NVIDIA

A Tesla car has several computers, at least one of them, (the MCU that controls that big screen) over time was NVIDIA based. Last year that MCU was already replaced with an Intel-based MCU for the newer models. There is one computer for the basic core functionality, and then another one for autopilot called the FSD, slowly advancing towards self-driving. 

That last part as far as we know it, is still NVIDIA based. However, Tesla just announced its own Self-driving AI Chip. A good while ago Tesla attracted Jim Keller from AMD to develop their own AI chip together with Peter Bannon.

The chip is 260mm 2 in size and contains 6 billion transistors divided into a CPU, GPU, memory, isp, h.265 video, and nna clusters. Most of these clusters are licensed from IP producers,  the neural network accelerator (nna) is completely self-designed. 

The new 14nm based self-developed chip is fabbed by Samsung and is capable of 144 tera operations per second (TOPS) (two chips each 72 TOPS) compared with Nvidia’s Drive Xavier’s theoretical performance of 21 TOPS. It features and two fully independent packages, each with their own with DRAM memory, flash storage chips, and power supplies. If opne fails, the failsafe is ta second one.

“The [packages] boot up and run their own operating systems, the two machines exchange … independent version of [driving] plans and make sure they’re the same, and assuming that they agree, they act and drive the car.”

That makes them highly redundant, Tesla CEO Elon Musk said. “Any part of this could fail, and the car will keep driving,” he added. 

The aforementioned FSD computers are built on a 14-nanometer FinFET CMOS process and measure 260 millimeters squared, with 6 billion transistors and 250 million logic gates. Their LPDDR4 RAM modules (running at 4,266 gigabits per second) boast a peak bandwidth of 68 GB/s, and the integrated image signal processors — which have 24-bit pipelines and support both advanced tone mapping and advanced noise reduction — can perform operations on up to 1 gigapixels per second.

There are two neural network accelerators in tow, both of which are clocked at 2GHz, have a 32MB of SRAM memory and 96 x 96 multiple and add arrays, and can process up to 1TB of data per second and perform 36 TOPS (72 TOPS total). Other on-chip components include a 1GHz graphics chip that supports both 32-bit and 64-bit floating point operations and a dozen Arm A72 64-bit CPUs clocked at 2.2GHz that deliver “two and a half times the performance of the current solution,”.

Musk said that Tesla will start offering retrofits to current Tesla owners who bought the ‘Full Self-Driving package’ in the next few months. Tesla mentions the chip is 21x more powerful for 0.8x the costs.