A look at Tesla’s Dojo supercomputer shared on Hot Chips 34

Tesla attended Hot Chips 34 and shared a wealth of astonishing information about their Dojo supercomputer and chip architecture.

The details shared about Dojo in Hot Chips 34 by Tesla’s Emil Talpes, who worked at AMD for about 17 years for Opteron processors, are only about the hardware and capabilities of the chips and Dojo as a whole. The achievement will be discussed at Tesla’s AI Day II on September 30.

According to Musk, the goal of Dojo is “to become really good at video training. We probably have the fourth or approaching the third most powerful computer center in the world for AI training. Our first goal with Dojo is to make it competitive and be more efficient and neural net training than a whole bunch of GPUs.”

Since Tesla needs a lot of computing power to process video data from the vehicles in its fleet, it has built a proprietary system-on-wafer solution. According to ServeTheHome“Each D1 die is integrated on a tile with 25 dies at 15kW. In addition to the 25 D1 dies, there are also 40 smaller I/O dies.”

All power and cooling are integrated directly on the Training Tile, which is capable of 10 TB/s on-tile bandwidth and 36 TB/s off-tile aggregate bandwidth. This architecture allows the tiles to scale with 9TB/s connections between them. They are also pluggable and do not require their own server.

“The defining goal of our application is scalability,” Talpes said at the end of the presentation. “We’ve downplayed several mechanisms that you find in typical CPUs, such as coherence, virtual memory, and global lookup directories, simply because those mechanisms don’t scale very well when we scale up to a very large system. Instead, we’ve relied on a very fast and very distributed network-wide SRAM storage. And this is supported by an order of magnitude higher speed of interconnection than what you find in a typical distributed system.”

The inside look at what Tesla’s building behind the scenes continues to prove how and why Tesla is at the forefront of artificial intelligence and neural network training. It gives the Tesla community an added sense of comfort knowing that Tesla will always have scalability and innovative technology at the forefront of everything the car company does.

See Anastasi In Tech’s recap of Tesla’s Hot Chip 34 presentation

Tesla continues to roll out Full Self Driving and has also increased the price of FSD. The system logs millions of miles and is constantly being improved as Elon Musk continues his quest to make driving safer, prevent injuries and save lives. However, there are still many skeptics of the program that has the potential to revolutionize transport as we know it.

Some of Tesla’s most prominent supporters are its fiercest critics, but Musk said at Tesla’s shareholder meeting in August that he welcomes honest evaluations. The YouTube channel for CyberOwners has a truthful, well-explained review. Mike Hoffman of CyberOwners clearly demonstrates and explains one of the main problems with Tesla’s ability to see its surroundings and navigate safely. The blind corner.

Video explanation

We have all encountered these situations while driving. Something obstructs our view and we have to lean over the steering wheel as we slowly creep forward to see if it is safe to continue. As Hoffman shows, the current camera setup has a large area it cannot detect. But there are solutions, some that would cost a lot of money and some that could be a simple camera swap.

Hoffman believes that Tesla may already be testing a new repeater camera. The repeater camera, more commonly known as the blind spot camera, is on the front fenders. These cameras record video on the side and rear of the vehicle. But Hoffman noticed something different in a video spotted by Ashok Elluswamy (below), Tesla’s director of the Autopilot program.

It appears that the program is collecting data from the repeater camera further ahead than the current camera view. There is also speculation that the video Elluswamy shows is from the b-pillar, the one between the front and rear doors. Hoffman says the camera looks much further ahead and is confident a new repeater camera will be used.

Alternative solutions to this blind corner dilemma include adding cameras to the front bumper and fog lights, or moving the b-pillar camera to the wing mirror housing. However, this would require new cameras and wiring, which would mean the computer would have to analyze more video streams. This action would be a significant cost to Tesla, and Hoffman believes a camera swap is much more likely.

Hoffman may be right. It wouldn’t be the first time Tesla has changed the repeater camera. When a software update in 2021 included the blind spot feature, some users found that the repeater camera had glare and was not much help. Tesla replaced these cameras free of charge.

Most of Tesla’s cameras are made to be quickly replaceable. Repeater cameras and b-pillar cameras can be changed in less than five minutes. That said, this change would require Tesla to replace potentially millions of cameras. Not a cheap solution, but perhaps the best for improving another aspect of a system at the forefront of technology.

Tesla continues to prove they make the safest cars on the road, with the Model Y receiving the highest safety rating from Euro NCAP and Australia’s ANCAP.

The Australasian New Car Assessment Program (ANCAP) is the latest car safety agency to give Tesla’s Model Y a top safety score. The Model Y only arrived in Australia a few months ago, despite being a popular vehicle in the US market for a few years now.

The Model Y also earned a top safety score from the National Highway Traffic Safety Administration (NHTSA) in the United States

ANCAP released its five-star safety rating for the Model Y. Tesla scored well in all categories, but excelled in adult occupant protection and safety assist.

Model Y achieved near-perfect results in the Adult Occupant Protection category with an excellent score of 97%. The highest score of any vehicle tested in this protocol. The result for the Adult Occupant Protection category is determined by a series of frontal, side and rear crash tests. In addition, the score includes several other safety factors such as rescue, recovery and post-crash safety.

Tesla will continue to iterate on the vehicle’s design and software to achieve even better safety results in the future.

In fact, Tesla CEO Elon Musk recently tweeted that Tesla has already further improved the Tesla Vision since these tests have taken place.

We have since improved it further

— Elon Musk (@elonmusk) 8 September 2022

However, there are always areas for improvement. Some of the categories Tesla can improve on include child protection and protection of vulnerable road users.

For child protection, the Model Y lost some safety points because it doesn’t support all types of child seats:

“Installation of typical child seats available in Australia and New Zealand showed that most child seats could be fitted in the rear seating positions, but the Type A capsule and one of the selected seat cushions could not be properly installed in the center rear seat position. ”

Another category where the Model Y fared worse was protecting vulnerable road users. ANCAP found that if the car hit a pedestrian, it would have difficulty at the bottom of the windscreen:

“The helmet on the Tesla Model Y provided GOOD or ADEQUATE protection to the head of a struck pedestrian over most of its surface, with POOR and POOR results recorded at the base of the windshield and on the rigid windshield pillars.”

Despite these criticisms, ANCAP was impressed with Tesla’s autonomous emergency braking system to protect road users:

“The Autonomous Emergency Braking (AEB) system is capable of detecting and reacting to pedestrians and cyclists. The AEB system showed GOOD performance in both daylight and low-light pedestrian test scenarios, with collisions avoided or mitigated in most scenarios, including reversing scenarios and some reverse (AEB Backover). In test scenarios for cyclists, the AEB system offered GOOD performance with maximum points scored.”

The agency was also impressed with Tesla’s autonomous driver assistance system to prevent/mitigate collisions with other vehicles:

“Tests of the AEB (Car-to-Car) system showed GOOD performance with collisions avoided or mitigated in all scenarios, including AEB Junction Assist, where the test vehicle can automatically brake to avoid a crash when turning across the path of an oncoming vehicle .”

Model Y earned a leading score of 98 percent in Euro NCAP’s Safety Assist category. This was thanks to Tesla Vision, the camera vision and neural network processing system that is standard in all Tesla vehicles in North America and Europe. With update 2022.24, Tesla has also begun switching some of its older radar-based vehicles to Tesla Vision.

Teslas are so safe because Tesla is constantly innovating and iterating on their vehicle designs.

Tesla recently added an airbag on the other side to prevent the front passengers from colliding with each other.