The Trend in Custom Chips
Since the development of the first microchips, the vast majority of products on the market use off-the-shelf parts as by far biggest benefit of using generic parts is that any function can be realised with a mix and match of different chips at affordable prices.
For example, a microcontroller can be combined with an ADC connected to a bus controller which itself is connected to an SPI flash memory. If the project requires a UART port, a transceiver can easily be added to the mix, and if more memory is needed, then another bus controller and memory module can be added. Of course, this isn’t the most efficient method for designing circuits, but it does allow engineers to focus on the function of a product instead of its inner workings. However, even in the early days of computing, there were edge cases whereby a dedicated chip provided better performance.
One example of such technology was the PLA (Programmable Logic Array) and ULA (Uncommitted Logic Array). PLAs are programmable logic devices that could be programmed by designers (sometimes after manufacture), while ULAs were typically mask-programmed gate arrays configured at the factory. These devices are a form of semi-custom silicon that allowed specific functions to be implemented either by designer programming or by factory mask programming, before being used in consumer products. The advantage of such devices is that they can provide high-speed logic which would otherwise be too difficult to integrate into a product. But while these devices were extremely useful, they did have their limitations.
Firstly, they often had fixed gate arrays or limited gate densities that constrained what they could implement. Secondly, they would have limited I/O counts, and thirdly, they were not truly custom silicon; designers could not change transistor geometries or create fully custom transistor-level layouts.
As semiconductors became more complex, it didn’t take long for some companies to start creating custom ICs. Companies such as Intel and AMD were well known for developing processors, and the rise of foundries like TSMC allowed fabless companies to have custom chips manufactured without owning fabs. This eventually led to the development of ASICs which allowed engineers to design custom semiconductor packages with unique transistor layouts. But it now seems that many large companies are moving straight into fully custom ICs or custom SoCs.
What has fuelled this trend is the use of technologies such as AI which have very specific power and design requirements. Unlike generic processors, AI processors need to be able to perform large-scale matrix multiplications efficiently using dedicated hardware, and this can be hard to achieve on generic processors. Furthermore, AI processors also require high memory bandwidth and often substantial on-chip memory, and this can be difficult to achieve when using generic parts. Thus, it makes more sense to go straight to a custom IC that is designed from the ground up with AI in mind.
Tesla Developing AI5 Chip
With all the challenges faced by Tesla over the past few years, one would think that Elon Musk would put his other projects on hold. But considering that Tesla is heavily dependent on self-driving capabilities for future vehicles, it makes sense that Musk is continuing to push Tesla’s AI development.
Now, Tesla is set to unveil its latest AI chip, the AI5, which will revolutionize Tesla's AI ambitions and position the company as a leader in the autonomous vehicle industry. Samsung will manufacture the AI5 chip using a 2nm process at a new Texas facility, expected to be fully operational by late 2026.
The AI5 chip, expected to be unveiled soon, marks a major shift in Tesla's in-house chip development strategy. Similar to the Apple M1 chip, the AI5 chip will feature custom neural cores and ARM CPUs, showcasing Tesla's expertise in integrating advanced AI technologies into automotive systems.
The inclusion of these cutting-edge chip technologies signals Tesla's commitment to enhancing the performance and efficiency of its self-driving systems. Tesla engineers have also streamlined the design, removing unnecessary components to focus the AI5 on AI task optimization.
According to Elon Musk, the AI5 chip will be a critical component in realizing Tesla's vision for fully autonomous vehicles. By utilising in-house chip design and manufacturing partnerships, Tesla aims to optimize its self-driving systems for real-world scenarios, setting a new standard for autonomous vehicle safety and performance.
The AI5 chip's integration into Tesla's upcoming Full Self-Driving (FSD) beta version and the Optimus robot prototype demonstrates the importance of customized hardware in achieving advanced levels of autonomous driving. With the AI5 chip powering FSD, Tesla is poised to take a significant step towards realizing the fully autonomous future it envisions. In addition to self-driving capabilities, the AI5 chip will also play a vital role in Tesla's robotic initiatives. Musk has hinted that the AI5 chip may even be used in data centres, offering up to ten times the throughput of previous models. This versatility of the AI5 chip showcases Tesla's ambition to apply AI technologies across various industries and applications.
Is the future all custom ICs?
For the time being, ICs are still mostly off-the-shelf with few businesses creating their own ICs from scratch, but this hasn’t stopped large businesses such as Apple and Google from doing just this. While these custom ICs often use the latest single-digit nanometre nodes, it may not be long before larger, older nodes (e.g., 250 nm-class and above) regain popularity for cost-sensitive applications.
If smaller transistors are not needed, then there is no point in paying extra money for a chip that uses them, and if larger transistors can provide the same performance at a reduced cost then it makes perfect sense to go with them. This is already a topic in the industry: rather than a one-way march to ever-smaller nodes, designers are recognizing that for many applications larger, mature process nodes remain attractive because they often cost less and can be more than sufficient.
So, if custom ICs become a big enough market, it may make sense for smaller businesses to invest in the tools and training needed to create custom ICs. However, other technologies such as chiplets could step up and do a better job by providing smaller basic building blocks that can be combined together.
For example, an IC package could have 16 slots for chiplets that allow for connecting any combination of CPU, memory, and GPU modules together. This would enable designers to create custom hardware with pre-made chiplets that are already optimised for each other.
Overall, custom ICs are certainly the future for hyper specific applications where traditional electronics simply cannot be used, but whether or not custom ICs become mainstream is yet to be seen.
You may also like
