Domestically, RISC-V is often referred to as a key path for my country's chip industry to change lanes, overtake, and become autonomous and controllable, and its development has also received widespread attention.

In the past few years, RISC-V has been mainly used in some simple MCUs, and shipments have been concentrated in some mid-to-low-end needs, such as printers, monitors, etc. However, now RISC-V is gradually moving into the field of high-performance computing, such as server CPUs, AI acceleration chips, automotive-grade chips, etc.

In fact, since 2018, the dynamics of RISC-V in cloud computing and supercomputing centers have continued to attract the attention of practitioners. However, just a few years later, the road to high performance of RISC-V is still long and difficult.

However, the industry generally believes that the RISC-V industry will see a turnaround in the past two years. It is not difficult to see from the news that the point where RISC-V will explode is getting closer and closer, not only because the industry scale is gradually growing, but also because new products are constantly being released and put into commercial use. This has already been seen in the server market. embodied.

First, in June last year, Suaneng came up with the RISC-V server chip SG2042; second, recently the French cloud server manufacturer Scaleway also launched a RISC-V server called Elastic Metal RV1. It is reported that the chip cores of both are from a domestic chip 910.

RISC-V server releases in the past two years

01 What kind of market is server chip?

Server chips are mainly processor chips designed for large-scale computing needs such as data centers, cloud computing, and enterprise-level applications. Currently, they are mainly divided into two markets: dedicated computing and general computing.

Among them, general-purpose servers can run many different types of software and services and are versatile and flexible. General server chips mainly include Intel Xeon, AMD EPYC, etc., which are used with general operating systems such as Linux and Windows Server. General-purpose servers are commonly used in enterprise data centers, cloud service providers, network servers and other scenarios, and are used to host various enterprise applications, websites, databases, storage, etc.

A dedicated server, on the other hand, is a server designed to perform a specific type of task or application. Dedicated servers are usually equipped with dedicated hardware and software, such as dedicated chips (such as GPU, TPU, etc.), specific accelerators, customized operating systems or application software, etc. Application scenarios for dedicated servers include high-performance computing (HPC), artificial intelligence training, video processing, blockchain nodes, game servers, etc.

In terms of chip architecture, current general-purpose servers are dominated by X86, and some end customers will use ARM. The Xinchuang market will choose self-developed architectures such as Loongson, while RISC-V is very rare.

According to industry insiders, the current server purchasers are mainly the three major domestic operators. In terms of procurement plans, the proportion of ARM servers has been set, but the procurement proportion has not yet been determined for RISC-V.

However, some cloud service providers said that many customers don't care whether they use ARM or x86, but they care about the system running faster or lower cost, or which server has more memory.

In the field of dedicated servers, the RISC-V architecture has long been favored by major manufacturers, such as Google, Amazon, and IBM in foreign countries, and Byte, Tencent, and Baidu in China. Each of them has been in its own cloud computing business since 2018. It has been using RISC-V server chips since 2008. However, it is understood that the market size is not large, and most of the above companies are self-researched and used to save costs. There is no commercialization logic for the RISC-V industry.

The current X86 chip players in the server market are mainly foreign Intel and AMD. Domestic Haiguang and Zhaoxin have also adopted the x86 architecture IP core licensing model. ARM players include Kunpeng and Feiteng, both of which have obtained permanent licenses for ARM V8. In addition, there are also new ARM entrants such as Yuxianwei, Hongjunwei, Qilingxin, etc.

It is understood that the core advantage of ARM's ability to squeeze into the server market is that ARM-based servers can support diverse and rapidly changing workloads, and are more scalable and cost-effective; secondly, they can provide different niche markets Higher customization and more flexible ecosystem; in addition, ARM architecture servers are relatively small in size and meet the needs of today's micro data centers.

So, does RISC-V have some differentiated advantages that can compete with X86 and ARM?

02 When will X86, ARM, and RISC-V achieve a three-legged alliance?

In the server field, the advantages of RISC-V compared to other architectures can be discussed in two aspects: technology and market.

Technically speaking, Academician Ni Guangnan mentioned at the recent Damo Academy Xuantie RISC-V Ecological Conference that compared with X86, RISC-V may surpass it in the following two aspects:

Today, the x86 architecture still monopolizes the server and desktop fields. This is because in the past few decades, the x86 architecture has formed a strong ecosystem in these fields, so that other architectures cannot replace it due to a relatively weak ecosystem. This will happen over time. Variety. On the other hand, the current x86 architecture is no longer pure CISC. In fact, it also uses hardware to convert certain CISC instructions into its internal RISC instructions for execution.

In the post-PC era, the competitiveness of the CPU architecture must not only be reflected in performance, but also in all aspects such as chip area, energy consumption, price, research and development cycle, ecological support, etc. By this measure, the x86 architecture not only has no advantage in emerging markets, but the total market volume in traditional fields will not grow much, and its share may even be cannibalized by RISG.

Compared with ARM, Academician Ni summarized the following advantages for RISC-V:

• Simple. The RISC-V architecture is much smaller than the proprietary ARM architecture. For example, the manual of ARMW8 is 2500 pages (similar to X86), while the manual of RISC-V is only 200 pages;

• Clean. RISC-V was launched 25 years later than ARM. The technology is new and has absorbed many lessons learned in the development of processor technology, such as avoiding features such as microarchitecture that depend on specific technologies.

• Modular. The overall architecture of RISC-V consists of a small and compact standard basic instruction set plus several standard extension instruction sets. Users can also add customized extended instruction sets.

• Customizable. A large opcode space is reserved for easy expansion; it is convenient for users to customize expansion instructions as needed.

From a market perspective, RISC-V, on the one hand, meets my country's demand for independent and controllable chips, and on the other hand, it can meet the customized needs of dedicated servers. Some practitioners said that in the future, RISC-V servers should start from the Xinchuang market (based on information security considerations), but in the future, there will be a need for customization because the CPU cannot improve performance.

In terms of price, the current licensing fees and royalties for RISC-V IP cores are only 1/3-1/4 of ARM. If the RISC-V solution is adopted, the costs of chip and server manufacturers can be greatly reduced.

However, industry insiders said that there are still several major challenges for RISC-V to enter the server market:

The first is the incompatibility of general software. The general server market requires the use of general operating systems and software, but currently many commercial software are not RISC-V compatible. Especially for a strong ecological market such as data centers, the work that needs to be completed includes a large amount of commercial software support, mutual collaboration between software, and the formulation of protocols/standards. The RISC-V Foundation has only just begun this part of the work.

The second is that adaptation is expensive. Server product finalization requires a lot of money (about 10 million), because it is not a simple assembly. There are many and complicated things that need to be considered, mainly stability. A server needs to run many cards, and there is no guarantee that different solutions will run smoothly. Easy, and very expensive during the pilot phase.

The third is that server adaptation is extremely difficult. Currently, most server companies buy ready-made boards because adapting the chip to the board requires a complicated process and a lot of time.

But despite the difficulties, RISC-V servers are now showing signs of light. Moreover, Academician Ni also mentioned at the event that RISC-V custom extension instructions are an effective way to implement DSA (domain-specific architecture), and introduced the opportunities for my country's new DSA servers based on the RISC-V architecture.

So, which players around the world are currently deploying RISC-V servers?

03 Which companies are deploying RISC-V servers?

At present, many domestic and foreign companies have entered the RISC-V server track, but some of them are mainly involved in AI acceleration chips in servers. IP companies include Alibaba Damo Academy, SiFive, Ventana, Tenstorrent, etc.; chip companies include Computing Power, Jindu Space-Time, Him Computing, Blue Core Computing Power, Microcore, Super Rui Technology, etc.

View more at EASELINK