Intel, Nvidia, AMD start a “full battle”

A series of major events one after another has provided more imagination space for the “competition” of the three giants Intel, NVIDIA, and AMD around the heterogeneous computing CPU+GPU+FPGA/DPU in the digital age, and has also become a new field of division in the future. annotation.

Intel is making a comeback in the field of independent GPU, and it is also making new innovations in the field of IPU, with the help of innovations in hardware, software, architecture and process, as well as the IDM2.0 strategy.

After AMD’s acquisition of Xilinx was settled, it made up for the shortcomings of FPGA. Not long ago, AMD announced the acquisition of cloud service provider Pensando for about 1.9 billion US dollars. At this point, AMD officially entered the DPU field and made up a key part of its data center blueprint. ring. Although NVIDIA was forced to “let go” by acquiring Arm, it has already used Arm-based CPUs as an important “supply”, and has completed DPUs through acquisitions, hoping to make a big difference in the era of heterogeneity.

The battle of the three giants has already penetrated into the hinterland, and the competition between Intel, NVIDIA, and AMD has shown a “comprehensive battle” situation.

The competition for “reboot” in the GPU field

In the field of heterogeneous computing, GPU can be said to be the “ammunition” that must be relied on.

As one of the biggest beneficiaries driven by the heterogeneous era and emerging applications, with the continuous improvement of computing power and AI performance requirements in the fields of servers, automobiles, artificial intelligence, and edge computing, GPUs rely on their own advantages in parallel processing and general computing. The advantages are advancing by leaps and bounds, and the market can continue to grow rapidly.

According to Verified Market Research, the global GPU market is valued at US$25.41 billion in 2020, and is expected to reach US$185.31 billion in 2027, with an average annual growth rate of 32.82%.

At present, GPUs are widely used in PCs, games, data centers, high-performance computing, smart cars and other fields. It is worth noting that games and PCs are its traditional main battlefields, while data centers, high-performance computing and smart cars will become new engines for GPU growth, and different applications have different demands on GPUs.

It is understood that the design ideas of game consoles focus on improving the experience, focusing on developers’ optimization of hardware such as CPU and GPU and software optimization such as the underlying API. The GPU of a PC needs to balance performance, scalability, and energy efficiency. There are mainly two types of integrated GPUs and independent GPUs. Most of the integrated GPUs have been integrated with the CPU as SoCs, while the independent GPUs mostly use the PCIe bus to communicate with the CPU in real time. From the perspective of high-performance computing and servers, GPUs have strict requirements for fast throughput of large data volumes, super stability, and long-term operation; automotive GPUs need to meet automotive regulatory certifications such as AEC-Q100, and support dedicated graphics APIs , and the future trend is that automotive CPU and GPU will form SoC, from distributed to centralized development.

After years of fierce battles, the global GPU has an oligopolistic pattern. Nvidia is the absolute hegemon, followed by AMD, but after Intel returns to the independent GPU battlefield, the original balance will be broken.

Through technological innovation, scenario expansion, extension mergers and acquisitions, and the continuous exploration of GPU general computing capabilities based on the CUDA software stack, NVIDIA has become a leader in the GPU field and leads the global GPU development. In fiscal 2022, Nvidia’s revenue hit a record $26.91 billion, up 61% year-over-year.

Looking at NVIDIA’s revenue structure, it can be found that, benefiting from the strong demand for NVIDIA’s Ampere architecture products, gaming has become the biggest driving force, and the data center market has the fastest growth rate, hitting a new high of US$10.61 billion; and although the automotive business has declined, it will continue to grow in the future. will continue to harvest. Its next layout is also full of firepower: a new generation of desktop GPUs and laptop GPUs have been launched; the next-generation GPU Hopper GH100 chip for data centers or more than 140 billion transistors will use TSMC’s 5nm node multi-chip module (MCM) design. And the next-generation autonomous driving chip Orin is planned to be used for mass production in 2022, and the computing power will reach 254TOPS. At present, it has won projects from several OEMs such as Weilai, Ideal, Volvo, and Mercedes-Benz.

After years of rapid progress, AMD has secured the second position in both the CPU and GPU markets. In terms of GPU layout, in 2022, AMD will further expand the graphics card market with new top, mid-range and entry-level GPUs, while equipped with new AMD Software support. In the field of data centers, AMD is also aggressive. Not long ago, it released the Instinct MI200 accelerator card based on GPU architecture, dedicated to HPC and AI acceleration. It uses the second-generation CDNA architecture (designed to optimize data center computing workloads), is the first multi-chip, the first GPU to support 128GB of HBM2E memory, and the first Exascale-class (exascale) GPU. It also introduced a new data center-oriented GPU, the next-generation Radeon Pro V620, designed to meet the growing demand for GPU acceleration for cloud applications, 3D workloads, and more.

Intel, which has a leading edge in the field of integrated GPUs such as PCs, has been aggressive and constantly improving since it announced its return to the independent GPU battlefield a few years ago. At the end of 2020, Intel debuted the Xe GPU architecture at its Architecture Day, the Xe microarchitecture to address needs ranging from integrated/entry graphics needs to data center and high performance computing. At the same time, Intel released its first data center server GPU, completing the comprehensive construction of the “CPU+GPU+FPGA” hybrid XPU architecture.

At Architecture Day 2021, Intel is launching two discrete GPUs. At the investment day held not long ago, Intel released two GPUs, one for the gaming field and one for the data center. Next, Intel announced that the data center GPU code-named ATS-M will be released in the third quarter, which integrates multiple Xe cores, AV1 hardware encoders, GDDR6 memory, ray tracing units, etc., and can provide 150 trillion operations per second. . Not only that, for the traditional PC field, Intel is also determined to win, and launched the Arc Ruixuan series graphics cards for laptop platforms and the first A3 series graphics card for desktops – Ruixuan A380 GPU. And, not just the A380, the Intel Sharp A5 series and A7 series with higher performance will also be available this summer.

In the GPU field where gunpowder smoke is everywhere, Intel, which is full of firepower, may challenge AMD and Nvidia in all directions.

Heterogeneous computing “hands-to-hand”

From a direct point of view, the heterogeneous “jigsaw” of the three giants Intel, Nvidia and AMD has been roughly formed.

Among these three giants, Intel’s heterogeneous combination is obviously more profound. In the past five years, Intel, which has established a “data-centric” transformation goal, has continued to enrich its layout in the data center field through mergers and acquisitions, including the acquisition of high-quality FPGA, eASIC, and ASIC companies, plus the development of independent GPU, IPU, Neuromorphic chips, quantum computing chips, and oneAPI, a unified programming software tool for research and development, provide a unified and simplified application development programming model for heterogeneous computing including CPU, GPU, FPGA and other accelerators, and realize a product portfolio covering multiple architectures.

Coupled with the recent massive expansion of the IDM2.0 strategy, as well as a series of actions to open x86 and join the RISC-V camp in a high-profile manner, Intel has more than one “trump card” in the era of isomerization, making it more comfortable.

From the perspective of AMD, its business has long focused on the two core areas of CPU and GPU, and FPGA is its biggest shortcoming. However, after AMD announced that it had completed the acquisition of Xilinx in an all-stock transaction, with Xilinx’s deep accumulation in the fields of FPGA, programmable SoC and ACAP, it provided AMD with the strengthening of horizontal cloud and edge computing capabilities. “Nutrition”. The merger of AMD and Xilinx will not only focus on improving its overall data center business competitiveness, but also gain more chips in the era of data center heterogeneity.

After Pensando was acquired by AMD, it means that AMD not only officially entered the DPU field, but also allowed AMD’s business to fully cover CPU, GPU, FPGA, and DPU, building a basically complete “puzzle” of computing power.

In order to fulfill its “GPU+CPU+DPU” route, NVIDIA first announced the acquisition of Arm in a high-profile manner, and then spent $6.9 billion to acquire Israeli network equipment maker Mellanox to supply DPU. Although the acquisition of Arm was ultimately “no accident”, it has invested heavily in CPU development, and officially launched its self-developed CPU for data center AI and high-performance computing applications at the GTC conference in 2021 – based on Arm Neoverse Architecture of the Grace chip. According to the agreement, NVIDIA has obtained ARM’s architecture license for nearly 20 years, and can develop ARM-based CPUs through ARM-licensed IP in the future.

For NVIDIA, the research and development of Grace CPU has far-reaching significance. Because GPU needs to be matched with CPU operation, this move will make it no longer limited in terms of CPU, and the self-reliance and self-reliance of CPU will also make its heterogeneous integration more obvious.

Faced with a comprehensive contest, the three giants also have different hidden worries.

According to industry analysts, AMD also needs time to digest and integrate GPU+CPU+DPU+FPGA to expand its ability to provide leading solutions for cloud, enterprise and edge customers; NVIDIA’s heavily relied GPU may face ASIC encroachment in the data center acceleration field in the future Intel is still a company whose genes belong to CPU, and the investment in GPU needs to match the growth of CPU, so it will be a huge challenge to deal with the development conflict between CPU and GPU. In addition, under the baton of IDM2.0, the focus of investment is inevitably tilted towards advanced manufacturing, and how to balance the innovation and integration of major XPU investment resources also needs to be carefully weighed.

It should be pointed out that with the determination of the Chiplet UCIe protocol, the design scale can be increased several times, for example, the CPU, GPU and DPU can all be expanded by N times in parallel; or to achieve vertical integration, CPU+GPU+DPU can be combined into a super-heterogeneous a single chip, or a combination of two.

Therefore, how to run different systems in parallel and how to efficiently and adaptively interact will become a new challenge for giants. Who can take the lead in this regard, who will magnify the future win.

Key factors affecting the pattern

After re-attaching to battle, the showdown of the three giants will also be full of firepower.

In addition to coping with the “xPU+” architectural innovation, ecological construction, and continuous tests of execution, it must be said that the process and packaging are the keys to turning ideas into actual products in order to achieve super-heterogeneous computing.

Let’s talk about the process first, and the related productivity factors.

Whether it is CPU, GPU, DPU, or FPGA, they are all pioneers of advanced technology. If you want to fight against a group of masters, the use of the most advanced technology is king.

Recent news shows that TSMC has difficulties with its 3nm process yield. If the 3nm yield problem continues, many customers may extend the use of the 5nm process node, thereby affecting the chip shipments of customers such as AMD, Intel, and Nvidia.

This makes supply bottlenecks caused by capacity shortages one of the obstacles they face. As Nvidia said in its earnings report, future supply constraints will remain a headwind given the global shortage of chip and wafer production capacity. Nvidia has reportedly prepaid TSMC about $1.64 billion in the third quarter of 2021 and will pay $1.79 billion in the first quarter of 2022, bringing the entire long-term order advance to $6.9 billion, much higher than what they paid earlier .

Intel’s advantage over Nvidia and AMD is its growing foundry business. Although Intel’s technology has not yet broken through 5nm in foundry, if it follows its technology roadmap, it will be on par with TSMC’s foundry level in 2025. Perhaps, at that time, Intel can fully support its own advanced process design, become more comfortable in the heterogeneous integration level of x86, Arm and RISC-V, and give priority to supply in terms of capacity guarantee. The deep meaning behind its IDM 2.0 strategy may be more profound than imagined .

In addition, heterogeneous computing cannot bypass heterogeneous integration and advanced packaging. The advancement of heterogeneous integration and advanced packaging technology makes it possible to build complex systems in a single package, which can quickly meet the power consumption, volume, and performance requirements of chips in heterogeneous computing systems.

At the level of advanced packaging, it seems that Intel, as a traditional IDM, seems to have more advantages, and AMD was originally an IDM, but later spun off the chip manufacturing business, but the company still has the genes of process and packaging. AMD has gained a head start over the past few years with its first-to-market chiplet and interconnect technology, building on the company’s next-generation packaging technology, 3D Stacked V-Cache. Here too, Xilinx can help AMD because Xilinx has built a range of high-performance packaging and interconnect technologies for its adaptive FPGA platform.

For NVIDIA, as a pure Fabless, it is slightly inferior to Intel and AMD in the process and packaging of heterogeneous integration, not only in the field of high-performance applications, but also more dependent on partners in terms of process and packaging.

In contrast, Intel has advanced in multiple ways, and has continued to advance in Co-EMIB, UCIe, Foveros, etc. Especially in the 3D packaging part, Intel has launched Foveros Direct, which realizes the transition to direct copper-to-copper bonding, and achieves bump pitches below 10 microns through HBI technology, allowing more than 10 times the interconnection between different chips. Density increase. And not long ago, the top-level accelerator card Ponte Vecchio developed for supercomputing, the number of integrated transistors exceeded 100 billion, using 5 different manufacturing processes, and encapsulating as many as 47 different units (Tile) inside, becoming a Foveros-based technology. The “integrator” of 3D stacked packaging technology and Co-EMIB connection technology.

According to data from consulting firm Yole Developpement, semiconductor manufacturers will spend about $11.9 billion in capital expenditures in advanced packaging in 2021. The agency said the advanced packaging market will be worth about $2.74 billion in 2021, and predicts that the market will achieve a compound annual growth rate of 19% by 2027, when the advanced packaging market will reach $7.87 billion per year.

From this point of view, the future competition will also be fully launched in architectural innovation, technology, packaging, etc. In these aspects, the three giants may need to cover everything.