In data centers, automotive, and other fields, an increasing number of companies are beginning to design their own SoCs.

As the cost of developing and manufacturing SoCs (Systems on a Chip) using cutting-edge manufacturing technologies continues to rise, more and more companies are opting to develop unique SoC designs to differentiate themselves from their competitors. This trend does not benefit standard CPU/SoC suppliers such as AMD, Intel, and Qualcomm. However, it has significant advantages for foundries and other industry participants, including EDA software developers, intellectual property (IP) providers, and contract chip design companies.

Several global trends are impacting the development of microelectronics. For instance: 5G, edge computing, software-defined vehicles (SDV), AI, cloud computing, high-performance computing (HPC), etc.

Applications such as artificial intelligence, cloud computing, and HPC may waste a certain amount of performance, but the main limiting factors are the power consumption and space of data centers, rather than the cost of hardware. Similarly, SDV and edge computing devices require higher computational performance. However, in addition to power consumption and space, they are also limited by cost (i.e., the number of chips used and the size of the chips).

Advertisement

Artificial intelligence, automotive, cloud computing, edge computing, and high-performance computing are completely different applications that can be addressed with widely used off-the-shelf chips. However, by precisely tailoring the chips to the workloads they run, it can more effectively achieve the required performance and power consumption levels from a cost and power perspective.

Arm's Vice President of Infrastructure Line of Business Product Solutions, Dermot O'Driscoll, stated: "We recognize that a one-size-fits-all approach can no longer meet computational needs, and more and more companies are considering custom SoC projects to obtain advanced computing capabilities."

Furthermore, intellectual property protection, functional differentiation, design flexibility, and the long-term availability of products are more important than ever. In response to this situation, many companies are developing SoCs in-house or ordering SoCs from contract chip development manufacturers.Ansys Chief Technology Officer Christophe Bianchi stated: "We will certainly see an increasing number of companies designing their own chips. Ansys is a leading supplier of analysis tool software and a developer of engineering simulation and EDA software. This shift is particularly evident in North America, but the EMEA (Europe, Middle East, and Africa) and APAC (Asia-Pacific) regions are now following suit."

Sondrel's founder and CEO, Graham Curren, said: "Having a custom SoC means having a custom SoC that integrates specific and differentiated technologies that are extremely difficult to replicate, which means having a unique solution that is hard to implement. It is deeply embedded in the semiconductor chip, protecting the company's intellectual property. Moreover, it is precisely tailored in terms of power consumption, performance, and area according to the task specifications, making it a fully compliant dedicated solution."

How Apple Benefits from Its Custom Chips

Apple is a great example of a newcomer that has developed its own SoC, initially for smartphones and then for personal computers, reaping huge profits.

Apple stated that it began asking Samsung Electronics to embed some of its own custom IP into its iPhone application processors from the late 2000s. Shortly thereafter, they established an in-house chip development team and developed custom CPU microarchitectures and accelerators to add value to products such as the iPhone, iPad, and Apple TV STB (set-top box). The company then started manufacturing SoCs for PCs, integrating various types of custom IP to achieve lower power consumption, longer battery life, and higher performance in demanding applications.

"The main reasons for designing custom SoCs include providing specialized acceleration, implementing unique system and packaging topologies, and optimizing costs through vertical integration," said O'Driscoll.

Apple is reaping significant benefits from its efforts with the M-series SoCs. In the first quarter of 2019, the company's market share in the PC market was 6.4% (shipments of 3,791,000 units), which increased to 9.3% (sales of 7,324,000 units) in the first quarter of 2022. Additionally, the company's R&D expenses for the fiscal year 2022 were $26.251 billion (higher than the $16.217 billion in the fiscal year 2019), with the majority allocated to semiconductor chip development.

In the past, semiconductor chip design was the domain of large, highly profitable manufacturers such as IBM, Cisco Systems, and Sun Microsystems (now Oracle). Subsequently, Apple demonstrated the utility of consumer-grade SoCs to the world, with companies like Huawei and Samsung following suit.

By the mid-2010s, internet giants had experienced significant growth and expansion of profits. It was no surprise that these companies began developing their own semiconductors.Now, automobile manufacturers are following in Apple's footsteps.

Dan Hutchenson, Vice President of TechInsights, a technology information service company, said: "In recent years, the number of manufacturers designing their own chips has grown exponentially. For instance, giants like IBM, Apple, Google, Meta, Amazon, Baidu, Tesla, and others are all developing their own chips. Starting around 2021, major automobile manufacturers began designing their own chips."

Leading Hyperscale Companies

Currently, Apple is the only major personal computer manufacturer developing its own custom processors. Given that PC manufacturers have significantly less control over Windows development than Apple does over MacOS, they are reluctant to develop their own SoCs.

Hyperscale companies differ from PC manufacturers in that they have near-absolute control over their own hardware and software. Therefore, it is natural for hyperscale vendors to develop their own custom SoCs.

"Hyperscale is the initial driving force behind custom chip development," says Bianchi, "The demand for artificial intelligence computing and cloud performance has driven this growth."

Google operates some of the world's most famous internet services, including Tensor Processing Units (TPUs) for artificial intelligence training and inference acceleration, Video Coding Units (VCUs) for YouTube video transcoding, and server SoCs, boasting a multifaceted custom data center chip strategy. Whether the rumored Google server SoC will outperform AMD's EPYC or Intel's Xeon remains to be seen.

In the meantime, Google's TPUs and VCUs have already provided the performance, scalability, functionality, and cost that the company needs to expand its AI/video services.

Karl Freund, founder and chief analyst at Cambrian-AI Research, said: "Cost and customization are the driving forces for these manufacturers to develop their own chips. One of the key applications we hope to support with chips is artificial intelligence workloads."EDA tools and IP giant Cadence's Vice President of Product Management for the Digital/Signoff Group, Kam Kittrell, stated: "Building super-scale chips is cheaper than buying chips in bulk, eliminating the middleman. Having valuable proprietary software allows for the delivery of software workloads with higher energy efficiency."

New players continue to emerge

SiPearl and Tenstorrent may be the most eye-catching new entrants, aiming to challenge industry giants in the development of custom chips. SiPearl provides Arm-based CPUs for Europe's exascale supercomputers. Tenstorrent is a developer and manufacturer of ultra-high-performance AI/HPC solutions based on the RISC-V instruction set architecture (ISA).

The two companies adopt different approaches.

SiPearl plans to commercialize its "Rhea" processor with up to 72 cores by 2024. It uses licensed off-the-shelf CPU cores "Arm Neoverse V1" and the design is handled by a semiconductor manufacturing contract manufacturer. SiPearl executives say they chose Arm cores for their optimal combination of energy efficiency, software compatibility, and cost.

SiPearl CEO Philippe Notton said: "We chose the Arm ISA because Arm has a proven server processor software ecosystem, which is a fundamental component of the market. We also chose the Arm Neoverse V1 cores because they are an ideal choice. Scalar/vector processing performance and energy efficiency (performance per 1W) are world-class for server processing and HPC, and the architecture can save millions of dollars in development costs and shorten development time by several years."

In addition, SiPearl also licenses standard blocks such as memory and PCIe controllers from IP suppliers, and obtains other IP from members of the European Processor Initiative (EPI).

"We are currently developing IP related to key performance/features, such as architectural implementation and memory hierarchy," Notton said, "We use standard blocks for PCIe and memory controllers, which are expensive and difficult to develop."

Synopsys states that SiPearl's approach is quite common in the industry.

John Koeter, Vice President of Marketing and Strategy for the Solutions Group at Synopsys, said: "For system manufacturers who are just getting into SoC design, it makes sense to acquire as much IP as possible, especially in interface IP and general-purpose processors where these companies will typically spend time and resources to differentiate themselves, such as interpreting algorithms and optimizing software workloads."Many companies distinguish themselves by designing specific blocks that are crucial to their systems. By doing so, you can create a USP (Unique Selling Proposition) and gain an advantage over competitors. This accounts for 10% of the required IP.

Tenstorrent aims to incorporate all possible performance, flexibility, and functionality into its chips and has decided to design SoCs (ultimately a multi-chip solution) based on the chosen custom RISC-V microarchitecture. This ensures an appropriate balance between performance and power consumption.

Tenstorrent's Vice President of Communications, Bob Grim, said: "Given the experience of Jim Keller and Chief CPU Architect Wei-Han Lien with Arm, we plan to use Arm to implement our high-performance CPU cores. We initially considered using Arm because of its strong performance. The software stack and a good compiler are essential, considering our focus on artificial intelligence, which led us to choose RISC-V."

The data format mentioned by Grim is BF16, which is widely used in AI applications. At the same time, Tenstorrent's RISC-V-based core for AI/ML workloads, Tensix, supports multiple data formats: BF4 and BF8, INT8, FP16, BF16, and even FP64. Currently, this combination is not available in commercial cores. If a company needs to add new formats to their future products, they can do so as is.

Grim stated: "Adopting RISC-V was a great decision because an open-source solution like RISC-V allows us to make the necessary changes when needed. If there are any bugs or issues, we can modify them without waiting for the licensor to release a modified or updated version. In fact, our experience with RISC-V for the Tensix core has shown that we can use RISC-V for high-performance CPU cores. We have confidence in using V."

However, despite using an open-source ISA, Tenstorrent has chosen to license interfaces and the like.

Tenstorrent's Vice President of Engineering, Stan Sokorac, said: "For developing industry-standard interfaces from scratch, such as PCIe, DDR, and Ethernet, we typically grant IP licenses."

Grim stated, "Development on the software side is progressing rapidly, and Tenstorrent plans to continue using RISC-V."

Grim added: "Once people turn to open source, they will not go back to proprietary solutions because the benefits of open source are so powerful." He said.GPUs Showcase Their Strength in the AI Field

Another approach to handling new artificial intelligence workloads is through the design of GPUs.

Jon Peddie, President of the American research firm Jon Peddie Research, stated: "The most common target market for new entrants is artificial intelligence, ray tracing, and gaming." In 2016, there were 5 companies producing GPUs for the x86 platform; now there are 17 manufacturers and several IP companies.

For Chinese companies like Biren Technology and MetaX, graphics processing is not the primary task; they aim to address the rapidly growing issues of artificial intelligence and high-performance computing first. In contrast, there is China's Moore Threads Technology. The company brings gaming GPUs to the market, not only as graphics cards for gaming but also as data center boards for remote gaming and AI workloads, clearly indicating the company's intention to capitalize on the AI trend.

The automotive industry is following Apple's path.

Behind this is the trend towards Software-Defined Vehicles (SDV). Cars are evolving into tightly interconnected software-defined platforms that can enhance their capabilities over time.

When cars perform hundreds of operations per second based on data from dozens of sensors, coordinating the numerous electronic components traditionally managed by on-board computers or ECUs (Electronic Control Units) becomes inefficient. Therefore, separating hardware from software and operating cars like traditional PCs is more practical. The latter can benefit from custom SoCs, and so can automobiles, which is why many car manufacturers are now designing their own chips.

Christophe Bianchi, Chief Technology Expert at simulation software provider Ansys, said: "The second wave of custom chip design comes from the automotive industry, which is not only a key component in implementing advanced ADAS (Advanced Driver Assistance Systems). This activity has made traditional car manufacturers realize that SDV is also key to increasing service revenue."

Interestingly, in terms of chip complexity, car manufacturers are quickly catching up with chip designs for hyperscale and AI/HPC. It is reported that many car manufacturers are already researching distributed multi-chip solutions."In the automotive industry, an increasing number of original equipment manufacturers are considering proprietary designs, which is undoubtedly driven by the growing complexity of ADAS applications for L3 and L4 level autonomous vehicles," Koeter analyzed.

Despite the rapid development of the automotive industry today, it is an extremely conservative industry because the equipment it produces (vehicles) is typically used for more than 10 years and has extremely high reliability.

Chip manufacturers with a good track record claim that their chips can last for many years, but no one knows what these companies will look like in 10 years. Therefore, one way for automobile manufacturers to ensure the lifespan of their key chips exceeds 10 years is to develop them in-house.

Bianchi stated: "Due to mergers and acquisitions, as well as spin-offs, the priorities and participants in the semiconductor industry can change rapidly, especially in the automotive industry, where it takes three to five years to develop and mature hardware and semiconductor platforms."