Source: Asia Vets Car Reference.

Before the nuclear bomb was dropped, Musk conducted a small 'test explosion.'

Just revealed$Tesla (TSLA.US)$Information about the next generation of in-car computing platforms:

It's not called HW5.0, it's called AI5 directly.

Energy consumption increased by 5 times, computing power increased by 10 times.

Why not call it HW5.0? What are the improvements?

HW is actually an abbreviation for hardware, meaning automatic driving system hardware literally.

This should be the internal engineering project code directly used as the product name by early Musk.

In the early stages of entrepreneurship, it was good to highlight Tesla's technology geek and engineering genes.

But now that Tesla is big and the global smart car industry is imitating and trying to surpass, it doesn't seem appropriate to be so casual.

So Musk announced that we changed our name, the new generation of automatic driving hardware platform, the name is called - AI5.

This belongs to the first principle of autonomous driving naming method.

The related information revealed by Musk also includes:

First, the computing power is increased by 10 times compared to the current HW4.0.

Secondly, the overall power consumption is increased by 4-5 times.

It will be launched in the second half of 2025.

There are a few points worth noting. In terms of computing power, Musk specifically mentioned that the computing power of the entire platform is about 10 times that of HW4.0.

In fact, Tesla has never explicitly published the specific data of HW4.0, only saying that it is 5 times that of HW3.0 (the 750TOPS data in domestic third-party information has no official confirmation). Therefore, based on known information and actual machine disassembly tests by foreign enthusiasts, it can be roughly inferred that the computing power of HW4.0 is between 300TOPS-500TOPS.

That is, a single FSD chip has more than 200 TOPS computing power.

By doing so, the AI5 platform's computing power may reach 3000-5000 TOPS.

This data is quite terrifying.

Currently, the top domestic players, such as Huawei, Xiaopeng, etc., which support NOA function for cities with and without maps, have a computing power platform of only 400-500 TOPS.

The 'core king' Nvidia Thor has a top configuration of only over 2000 TOPS, and many opinions in the industry believe that mass-produced cars do not need such excessive computing power.

Is Musk trying to boost public confidence in Tesla by launching satellites, or has he really made revolutionary changes to the paradigm of autonomous driving technology? This is a big suspense and highlight.

The second thing worth noting is energy consumption. Musk mentioned the fifth-generation hardware platform, which has an overall energy consumption of 4-5 times higher than that of HW4.0 and proves the further software-hardware integration and more reasonable energy consumption of Tesla's autonomous driving.

It should also be added that one of the differences between a vehicle computing chip and a smartphone computing chip is that it can indulge in "richness" without having to consider energy consumption and size as important indicators. The heat dissipation conditions of the vehicle computing platform are naturally much better than those of PCs and smartphones.

After all, the battery capacity of electric vehicles is much larger than that of mobile phones. Even if the chip consumes more power, it is only a small part compared to the electric machine, air conditioner, etc.

Also, it is worth mentioning that Musk said that AI5 is the "latest generation computing platform" for Tesla, but is Tesla's business only about cars?

So this is also the basis for most netizens' speculation: AI5 is likely not specifically prepared for FSD, and such high computing power may match the Tesla Robotaxi and humanoid robot Optimus to be launched in August this year, as well as future Tesla models that support L4, L5.

Perhaps at that time, Tesla's products will really eliminate the driver's compartment.

But there is also a sharp question raised by users:

When HW3.0 was released, Musk guaranteed it could achieve unsupervised FSD function, but obviously it was delayed. Even today, HW4.0 still cannot.

Can AI5 definitely achieve it?

What changes will AI5 bring?

Now the only confirmed information is that AI5 is manufactured by Samsung using the 4nm process.

Initially, Tesla had said it would 100% be manufactured by TSMC, but after Musk met with Samsung executives, they changed their initial decision.

It's said that Samsung offered Musk an "unmissable" preferential price, and the yield rate of advanced processes reached over 70%, which is comparable to TSMC's.

So it can be basically confirmed that AI5 will still use the Exynos-IP core based on Samsung's design. Exynos-IP is Samsung's own IP based on the ARM architecture, and Google's mobile phone has also used it.

In 2019, Samsung stopped related work because of cooperation with Qualcomm, but Exynos-IP was designed in a very advanced manner. It is basically equivalent to the design of ARM Cortex X3, which is the flagship of the current ARM Cortex X series. Therefore, HW4.0 designed after 2020, as exposed by Greentheonly, a foreign Tesla breaking news expert, still uses Exynos-IP.

AI5 is likely to continue this route.

Of course, times have changed, and the technical paradigm of autonomous driving algorithms has undergone a profound transformation, from the initial CNN-based modular network structure to the end-to-end integrated network dominated by Transformer.

The outside world is typically more concerned with the NPU computing power of autonomous driving chips, and this is not wrong. In the era of large models, stronger AI computing power is certainly necessary.

However, for autonomous driving tasks, it is not enough for just NPU to be more powerful. Even if the AI processor is fast and has high computing power, if 90% of the time is spent waiting for memory to transfer data, it is useless.

In the era of CNN, an external CPU could work well with an AI processor, but by the time of Transformer, the CPU had become a bottleneck and internal integration became the optimal choice.

In fact, HW4.0 added a CPU to deal with Transformer, which coincidentally follows the same principle as Tesla's supercomputer chip Dojo's D1 architecture.

This also explains why, while Musk was teasing AI5, he specifically emphasized that HW4.0 was not being retired directly, but was being used to build a training cluster (and serving simultaneously with Nvidia's A100).

Therefore, we can also infer what Tesla's arrangement and impact will be for the next generation computing platform AI5.

The ultra-high computing power of AI5 is most likely the result of multiple next-generation 4nm FSD chips combined, rather than a single or dual chip implementation.

If it is really a single/dual-chip with thousands of TOPS, the cost will be almost impossible to mass-produce, which is not a viable option for either the automotive industry or the technology industry.

This also means that AI5 is likely to be a platform solution, allowing Musk to determine the specifications based on different tasks, scenarios, and product cost requirements.

The most complex humanoid robots may require thousands of TOPS of computing power, while Robotaxi may require slightly less. Correspondingly, the production vehicles by Tesla may require even less.

As for how much computing power is needed for fully unsupervised FSD on the vehicle side, even Musk himself may not be clear at this point and may still be exploring it.

The 'end-to-end' most reasonable and efficient mode may be to model all AI and make it into a large black box, or it may be a gradual integration with certain conditions. There is no consensus in the industry.

It is difficult to judge how large the computing power of AI5 will be reflected in Tesla's production vehicles, but it is certainly greater than the current four to five hundred TOPS.

Therefore, the most direct impact and influence brought by AI5 is to open up the computing power arms race for intelligent cars and autonomous driving.

In this competition, all industry players will gradually clarify the basic prerequisites and thresholds for end-to-end autonomous driving.

For L4, AI5's exploration and trial on Tesla's Robotaxi is as important as production vehicles, in that it creates a template for Robotaxi: how much computing power, what kind of configuration, and at what cost.

More importantly, Tesla is leading by example as this AGI player, demonstrating what kind of technical system is needed, how to build a universal platform in the backend, and how to choose cloud-terminal architecture...

Editor/Lambor