Domestic GPUs, this wave of momentum is a bit fierce.
First of all, at the end of last year, Core City released the "Fenghua No. 1" as a reference for the high-performance desktop/server graphics card market.
Moore Thread mass-produced a full-featured domestic GPU in just 18 months. Prior to this, Moore had released two products, namely Moore 1 and Moore 2, the former using Intel's dedicated CPU, and the latter using Intel's general-purpose processor. Biling Technology was also announced after the successful lighting of the first general-purpose GPU chip.
Such speed and achievement is even amazing:
Is this the rhythm of spending the night under the old yellow city walls?
The first things to consider are the product itself, performance, price, and time to buy.
The industry behind this suddenly caught fire, and the reason was: "How did this happen?" "Isn't that called a patent?" "That's our strength too." "Patents are a good thing, but we have to have our own core competitiveness." "How can you be so stupid? "How can patents be exempted?"
In the heated discussion, the term "IP authorization" has also surfaced.
For example, Imagination, a chip designer that has been repeatedly mentioned in this round of discussions, is one of them.
According to reports, this wave of new products of GPU head start-ups is inseparable from many of its factors. (.) Among these enterprises, "cloud computing + big data" is perhaps one of the hottest topics. "Cloud computing" refers to the combination of computers and the Internet. "Big data" refers to the collection of massive information resources. (.) (.) (.) (.)
A Chinese-funded chip design company
Imagination Technologies, like the blockbuster ARM, is a UK-based chip design company that makes money by licensing chip IP to other manufacturers and licensing fees.
At that time, the company was known as "Apple's royal IP licensor".
From the first generation to the iPhone7Plus, the graphics processor portion of the A-series chips is built on Imagination's Power VRIP core design.
According to statistics, in 2010, more than 215 million chips containing Power VR technology were sold worldwide.
In the mobile GPUIP licensing market in 2012, Imagination accounted for nearly 50% of the market.
In addition, according to the Financial Times, Intel and Apple both owned a large stake in the chip design company, so that the two sides were caught in a tug-of-war.
In 2016, it was also reported that Apple and Imagination had "senior negotiations" over the acquisition. As soon as the news came out, it immediately triggered a heated discussion among people. Both Apple and Google have said the two sides will complete the deal by the end of 2018, and the two sides also plan to sign an agreement in early 2019. So why did Apple choose to acquire Imagination? But in the end, the acquisition failed.
2017 ushered in a big turning point.
As early as April 2017, Apple announced a "breakup" with Imagination and committed to independently developing GPUs.
As a result, Imagination's stock price plummeted by more than 70% – at that time, according to the Financial Times, more than half of Imagination's revenue came from Apple.
Imagination's GPU business has been hit ARMMali.
In order to expand the market share of Mali GPUs, ARM has adopted a "buy CPU, send GPU" strategy, if you buy CPU core license, you can get GPU core discounts.
According to a number of media reports, although Imagination is in a leading position in mobile GPU technology, ARM's "buy one get one free" strategy has led many mobile phone chip manufacturers to change their stance.
Ultimately, Chinese-backed private equity funds upgraded Imagination: Canyon Bridge, funded by the Guoxin Fund, acquired Imagination Technologies for £550 million.
Notably, Imagination also has a CPU chip architecture MIPS in its pocket, but sells the business separately while holding hands with Canyon Bridge.
While today's mobile GPU No. 1 has been snapped up by ARM, Imagination's momentum has grown in recent years.
In particular, its investment in mobile light-chasing technology is now beginning to bear fruit.
Although the official did not specify which domestic chip manufacturer Imagine Excore cooperated, game companies such as Perfect World Technology Perfect World, Tencent Games and Net Ease have all begun to develop "Power VRPhoton, an advanced optical light-chasing architecture in the field of mobile GPU."
Apple also renewed a new multi-year licensing agreement with Imagination in January 2020.
How IP licensing works
That being said, what role do IP vendors like Imagination play in GPU design and development?
The first thing to explain is that the IP mentioned in the chip field is Intellectual Property.
Validated, reusable integrated circuit design modules are also known as IP cores.
In the early days, chip companies would modularly reuse the circuits they developed to enable certain functions, namely "don't repeat the wheel."
With the increasing scale of integrated circuits and the increasing complexity of chip functions, the market rhythm of various electronic products is also accelerating.
It is difficult for chip companies to design all the modules independently, and the use of mature IP cores is an inevitable choice to reduce costs and shorten the design cycle, and it is also an industry practice.
Suppliers specializing in IP verification and authorization have also left the industrial chain.
In this space, the biggest contenders are ARM, ARMCortex-A710 and X2, which are used in the CPU parts of mobile chips, such as Qualcomm and United Ex, the latter of which are microarchitectures licensed in the form of IP cores.
According to the form of authorization, IP cores are divided into soft core, solid core and hard core.
The soft core is a circuit module described in a hardware description language such as Verilog/VHDL.
Solid core is based on the soft core to complete the design and verification of the gate circuit netlist, sent to the Gate-level Netlist.
The hardcore further completes the physical layout, including simulation verification, and provides the final product mask layout during the design phase.
From soft to hard, the less work needs to be done to buy an IP core, the less room for adjustment.
However, even though IP cores can be purchased from third parties, the design of a complete GPU is still not easy.
Take Nvidia's recent release of Hopper, for example. Hopper is a full-featured, full-featured GPU architecture that includes a separate stream processor for image rendering, a tensor unit for data computation, rasterization units (ROPs), ray tracing units, and more.
There are also peripheral video memory, caching, video decoding, power management, PCIe bus interface, etc.
△ Nvidia Hopper architecture
According to IBS, the average number of IP cores that can be integrated in a single chip under the 7nm process is 178 and will increase to 218 5nm.
It's still difficult to integrate so many IP cores together to ensure efficient collaboration between them.
Soft IP needs to be hardened, integrating interfaces between different IPs, as well as overall chip layout design, overall verification, and testing. (.) This article will go into detail about these aspects and give a complete solution. The scheme has been applied in practice. Practice has proven to be feasible and effective. It meets system requirements well. (.) (.) (.) (.)
Any oversight can lead to the failure of the tape-out, the need for redoing, hundreds of millions of dollars in tape-out costs, and more upfront investment.
In the GPUIP licensing market, ARM, Imagination and VeriSilicon currently account for the top 3% of the world's shares.
The styles and situations of all three are different.
ARM's Mali series GPUs are featured products for mobile terminals and have been selected by MediaTek, Samsung and Huawei HiSilicon. At this year's CES show, MediaTek released the first processor using the technology, the S3C2440, while Samsung announced that it will work with HiSilicon to develop a mobile computing platform that supports intel cores. But HiSilicon's future is unclear, and Samsung has begun to develop its own IP under the AMD architecture license.
Imagination is more focused on GPUs, and Power VR outperformed Qualcomm and ARM in the early days, but with a single revenue stream, Imagination will be hit hard by the losses of big customer Apple, who hopes to break into the desktop, car computer and high-performance computing markets.
Silicon, a well-known ip provider in China, also provides one-stop chip design services and obtains GPU services from Vivante. Globally, VeriSilicon Microelectronics has strong technical strength and rich product experience. At present, it is the first chip IP licensing service provider in China and the seventh in the world.
Source: Core Original Annual Report 2021
Domestic GPU companies have recently taken the difficult first step of using IP licensing for multiple flowering, successful tape-outs, or mass production.
The IP licensing model not only reduces investment and accelerates returns, but also allows for rapid accumulation of experience and talent.
After accumulation, it is possible to embark on the road of architecture authorization + IP self-development like Apple and Samsung.
In fact, even Apple's self-developed IP has not been smooth sailing, and after three years of spin-off, the company returned to work with Imagination, which is said to have patent reasons.
IP nuclear suppliers have built moats with technological advantages and patent rights, and chip design companies must pay more to bypass these moats.
Imaging Imagination GPU Qualcomm is the leader, and if Apple sticks to its own research, it may bypass existing patents and come up with new methods without keeping up with product updates.
Finally, the success of GPUs lies not only in the hardware, but also in the support of programming interfaces and software.
In addition to the powerful hardware performance CUDA's computing CUDA computing platform ecosystem is an important reason why companies such as NVIDIA can dominate the GPU market.
AMD's ROCm open computing platform is unlikely to shake CUDA's dominance in the short term.
Domestic GPU startups, such as Moore Thread and Wallin Tech, still have a long way to go.
Domestic GPUs, why now?
The last batch of self-developed domestic chip entrepreneurship coincided with the AI boom, mainly focusing on data computing, that is, various types of DPU and NPU.
15 years ago, with the birth of a number of AI chip unicorns such as Horizon, Cambrian, and Flinthara, large enterprises such as BAT Huawei have followed suit.
Since the second half of 2020, fully functional GPUs, including graphic rendering, have become a new topic in the capital market.
Biling Technology, Moore Thread and many other established financing companies, Core Technology Company, Mega Core Technology and many other established chip companies have also announced independent graphics card projects near here.
A number of events from this period may have had an impact.
First of all, it should be noted that in 2020, the mainland will comprehensively promote the xinchuang industry, including chips, servers domestic substitution and independent controllability.
Chipping Technology "Xinchuang" was listed as one of the main application directions of Fenghua No. 1, and reached ecological cooperation with the domestic operating system Tongxin Unicom to jointly complete the transformation of domestic GPUs and domestic OS.
The official website of Xindong Technology
In addition, due to the global chip shortage and mining, the price of GPUs themselves has soared since 2020.
The main origin price of Ethereum GPU rose 4.8 times from the previous month, up 1.8% year-on-year; as the market demand for Ethereum continues to increase, ChipPeel has also begun to get involved in this field. As the world's largest IP hardware manufacturer and software developer, Chiptron technology has rich experience in IP hardware development and strong hardware and software solution capabilities. Prior to this, one of the businesses of Chippe Technology was the production of ASIC mining machines for the mining of bitcoin.
The official website of Xindong Technology
Finally, the explosion of metaversms, digital people, and AR/VR has also pushed the need for 3D graphics computing to new heights.
The application direction disclosed by Biling Technology mainly focuses on automatic driving and metacosmity, and Moore thread also proposed "metacomputing".
Moore Thread official website
Because of this, in less than two years, the start-up projects or projects in the middle of 2020 will produce the results of the domestic GPU or the speed of tape-out.
One of the core technologies, Fenghua No. 1, explicitly uses Imagination's IMGB series of GPUIP, while the others do not provide details.