Source: Content by Semiconductor Industry Watch (ID: icbank) original, author: Chang Qiu, thank you.
In recent years, the importance of GPU in the industry has become more prominent, whether in high-performance computing, or in the consumer-grade field, its stickiness to users is getting stronger and stronger, Nvidia's popularity is due to its core GPU technology and products, in this case, the traditional giant Intel can not sit still, originally only in the consumer market to produce integrated GPU graphics, changes in market demand made Intel began to set up an independent GPU research and development team, and invested more and more resources to cope with the competition between Nvidia and AMD. Especially in the field of high-performance computing.
In the field of high-performance applications, the power consumption and cost controllable requirements of the GPU are getting higher and higher, which puts forward higher requirements for related technologies, including chip design methods, EDA tools, process processes, and packaging technology, in order to achieve an effective balance between high performance and power consumption and cost, the above technical links are indispensable, and with the gradual "failure" of Moore's Law, the importance of advanced packaging technology is becoming more and more prominent, and Intel, AMD and NVIDIA, the three giants have seen the importance of this link, And constantly strengthen research and development efforts. Especially recently, these three companies have invariably disclosed important information in MCM (multi-chip module).
MCM hits the GPU
MCM is to solve the problem of low integration of a single chip and imperfect functions, it will be a number of high integration, high performance, high reliability die, in the high-density multi-layer interconnection substrate with SMD technology to form a variety of electronic module systems, the formation of multi-chip modules. MCM has the following characteristics: the package delay time is reduced, it is easy to achieve high-speed module speed, the package size and weight of the whole machine/module are reduced, and the system reliability is greatly improved.
In the past, MCM was mainly used for CPU and storage devices, especially in the CPU field, such as the early IBM Power4 dual-core processor, that is, the MCM formed by 4 dual-core Power4 and additional L3 caches, as well as Intel's Pentium D (R&D codename: Presler), Xeon, and AMD's Zen 2 architecture Ryzen (core codename: Matisse), EPYC processor, etc. They are all typical representatives of the application of MCM.
In recent years, under the leadership of AMD, MCM packaging technology has begun to move towards GPUs. This is mainly because traditional graphics cards are PCB line cards with multiple GPUs that require a Crossfire or SLI bridge that connects two separate graphics cards. Traditional SLI and CrossFire require a PCIe bus to exchange data, textures, synchronization, and so on. Since rendering times between GPUs create synchronization problems, in many cases, traditional dual GPU graphics cards, i.e. two chips on a single PCB, are interconnected by it, each with its own VRAM. SLI or CrossFire consumes a lot of energy, and cooling is a challenge that has plagued engineers for a long time.
The MCM GPU is a separate package with an on-board bridge that replaces the traditional Crossfire or SLI bridge between two discrete graphics cards.
In high-performance computing applications, the advantages of this MCM GPU are clear, and it is worth spending more time and effort on solving software problems in terms of packaging and interconnection to cope with the higher complexity of MCM designs. At present, MCM GPUs are mainly used in data center and cloud computing applications. As technology continues to mature and PC application performance improves, its applications in the field of consumer electronics will also emerge.
The Big Three are making a lot of efforts
The first to introduce MCM packaging technology to GPUs was AMD. In 2020, the company separated the GPU architecture of the game card and the professional card, the architecture of the game card is RDNA, and the architecture of the professional card is called CDNA, and the first product is the Instinct MI100 series. In 2021, AMD's Q2 financial report confirmed that the CDNA 2 GPU has been shipped to customers, and its GPU core code name is Aldebaran, which became AMD's first product in MCM packaging, which is prepared for data centers. On the PC side, after the introduction of the next-generation RDNA 3 architecture in 2022, MCM-based consumer Radeon GPUs will also emerge.
Manufacturing a multi-chip computing GPU is similar to making a multi-core MCM CPU, such as a Ryzen 5000 or Threadripper processor. First, bringing the chip closer together can improve computational efficiency. AMD's Infinity architecture ensures high-performance interconnects, which are expected to make the efficiency of both chips close to one. Second, it is easier to mass-produce multiple small chips using advanced process techniques than large chips, because small chips are generally less defective and therefore have better yields than large chips.
The other day, on the 2021 earnings call, AMD confirmed that there will be several important product launches this year, including GPUs based on the RDNA 3 architecture, that is, the Radeon RX 7000. At present, the latest graphics card in the series will have three GPUs, namely Navi 31, Navi 32 and Navi 33, of which Navi 31 and Navi 32 will use MCM packages. There have been rumors that the Infinity Cache of Navi 31 and Navi 32 will use the design of the 3D stack and will be added to the MCD chip separately, similar to the principle of using 3D V-Cache on the Zen 3 architecture, and the performance will be greatly improved.
Since the Navi 31 and Navi 32 use MCM packages, AMD will use two different processes, with GPUs using TSMC's 5nm process and caching I/O chips using TSMC's 6nm process.
Nvidia is also following up on the MCM package GPU.
In 2017, Nvidia demonstrated a design built from four small chips that not only improve performance, but also help increase yield (smaller chip yields increase), and also allow more computing resources to be pooled together. This multi-chip design also helps to improve power supply efficiency and has better heat dissipation.
Recently, Nvidia researchers published a technical article outlining the company's exploration of MCM, Nvidia's current practice on MCM packaging GPUs called "Composable On Package GPU" (COPA), the team described the advantages of COPA GPUs, especially the ability to adapt to various types of deep learning workloads.
As traditional converged GPU solutions are rapidly becoming less practical, researchers came up with the idea of COPA-GPU. Converged GPU solutions rely on an architecture of traditional chips, complemented by a combination of specialized hardware such as high-bandwidth memory (HBM), tensor cores/Matrix Cores, and ray tracing (RT).
This type of hardware may be well suited for certain tasks, but it is inefficient in other situations. Unlike current monolithic GPU designs that combine all specific actors and caches into a single package, the COPA-GPU architecture has the ability to mix/match multiple blocks of hardware. This allows it to better adapt to today's dynamic workloads that high-performance computing can only present, as well as deep learning (DL) environments.
This consolidation is more adaptable to multiple types of workloads, resulting in higher levels of GPU reuse. What's more, for data scientists, it gives them a better position to tap existing resources to push the boundaries of potential.
For the data center and consumer markets, NVIDIA will launch GPUs based on The Hopper architecture and the Ada Lovelace architecture, respectively. It is reported that the company will only use MCM technology on Hopper architecture GPUs, and Ada Lovelace architecture GPUs will still retain the traditional packaging design, and will not introduce MCM multi-chip packaging to consumer GPUs like AMD's Navi 31 based on RDNA 3 architecture.
Recently, it has been reported that the number of transistors based on the Hopper architecture GH100 will reach 140 billion, which is almost 2.5 times that of the current GA100 (54.2 billion) based on the Ampere architecture or the AMD Instinct MI200 series (58 billion) based on the CDNA 2 architecture. The chip size of the GH100 is said to be close to 900mm, which is smaller than the previously rumored 1000mm, but it is larger than the GA100 (862mm) and Instinct MI200 series (about 790mm). The GH100 is rumored to be configured with a total of 288 SM, providing three times the performance of the A100 compute card.
It is reported that as Nvidia's first GPU based on MCM technology, Hopper architecture products will use TSMC's 5nm process process to support HBM2e and other connectivity features, and are expected to be unveiled in mid-2022, and competitors will be Intel's Xe-HP architecture GPUs and AMD's CDNA 2 architecture products.
However, the above statement has not been officially confirmed, NVIDIA will hold a GTC 2022 conference on March 21 this year, at which time, the Hopper architecture and the corresponding acceleration card plan may be announced.
As a latecomer to independent GPUs, Intel has also been moving frequently recently.
Recently, Intel announced a new patent describing how multiple compute modules work together to perform image rendering, which means that Intel GPUs will use MCM packaging technology to greatly improve operational performance.
Intel's CPUs for data centers and supercomputers, the Ponte Vecchio, have been designed using a multi-chip and MCM packaging technology. In the new patent, Intel proposes a GPU image rendering solution that integrates multiple chips into the same unit to address issues such as manufacturing and power consumption while optimizing scalability and interconnectivity.
At present, this kind of image rendering problem will be solved by algorithms such as Alternate Frame Rendering (AFR) or Scissor Frame Rendering (SFR), but Intel is a checkerboard format rendering that integrates computing modules, and there are distributed operations, making multi-chip design GPUs have higher computational efficiency. Although Intel did not elaborate on the architectural level details, it can be expected that intel Arc brand graphics with MCM packaging technology GPUs should only be a matter of time.
epilogue
In terms of GPU research and development, Intel, AMD and Nvidia appear to be more and more "synchronized", especially in the process process and packaging technology, the process relies on TSMC, the package values MCM, in these two aspects of the original leading AMD, its advantages are getting smaller and smaller, especially in MCM, Nvidia and Intel are developing faster and faster, not only in the field of high-performance computing, in the consumer market, although AMD first applied MCM technology to PCs, but Nvidia and Intel are also accelerating progress, I believe that there will be corresponding programs launched soon.
MCM package GPUs began to enter the era of the Three Kingdoms.