Shulou(Shulou.com)06/02 Report--

The semiconductor chip industry, suddenly, brushed the screen…

Trump used all the power of the United States to lift Huawei. This wave of harassment is really eye-catching.

However, when the two sides argue about "taking away my chips,""focusing on our own, beautiful alone," perhaps we can focus on the technology itself to talk about the necessary and important chip industry layout.

The first one is the heterogeneous calculation of the same "spare tire to normal."

We know that the semiconductor industry has long focused on a few chips. Whether X86, ARM, RISC, the computing unit inside a CPU is the same architecture. The so-called heterogeneity is to combine CPU, DSP, GPU, ASIC, FPGA and other computing units with different process architectures, different instruction sets and different functions to form a mixed computing system.

Heterogeneous computing technology was born in the 1980s, but it has only begun to show its edge in the industry in the past two years, and quickly replaced general purpose CPU, standing on the industry "C position."

For example, Intel's latest AI platform includes a series of different processing cores such as CPU, GPU, DSP, NNP, FPGA, etc. Nvidia's robot platform Jetson Xavier also contains six processors, GPU/CPU/NPU/NVDLA, etc./none less.

Many cloud computing service providers have also upgraded their heterogeneous computing solutions. For example, Huawei's database GaussDB product just released last week uses X86, ARM, GPU, NPU and other computing power to perform calculations.

Smartphone SoC also began to add accelerated DSP, graphics processing unit NPU, etc. in addition to traditional CPU/GPU/ISP/baseband chips.

Then the question comes. In the field of supercomputing, isomorphic computing has replaced isomorphic computing and become the object of competition among chip manufacturers. What is heterogeneous computing based on? Behind the semiconductor industry's collective embrace of heterogeneous computing, are there old problems and new opportunities buried?

WHY: Heterogeneous computing epistatic history

First to answer a question, why is the CPU used well, and everyone suddenly collectively played the idea of heterogeneous computing?

The most direct reason is the rapid rise of computation-intensive fields, facing the explosive growth of computing demand, making a single chip more and more powerless.

In recent years, semiconductor technology has reached the physical limit of vertical speed increase, and processor performance can no longer follow Moore's Law (doubling every 18 months).

Intel extended its R & D cycle from two years to three years in 2016. Due to the limitation of CPU parallel computing power, supercomputers often have to work in parallel with tens of thousands of processors.

Another key impact is the increasing importance of AI in computing scenarios. in particular Mobile device, AI is appearing in application in a variety of modes. In addition to basic communication capabilities such as phone calls and text messages, it is also necessary to process various information such as pictures, entertainment games, high-definition cameras, and provide personalized intelligent push, prediction, etc. These computing requirements have far exceeded the capabilities of traditional CPU processors.

In response, GPU chip manufacturer NVIDIA directly put forward a new slogan for the semiconductor industry in 2017:"Moore's Law is dead, long live artificial intelligence!" "(Moore's Law is dead, long live AI.)。

When a single chip cannot meet the computing power requirements of high-performance computing, heterogeneous computing is selected by the times.

A brief summary of the core advantages of heterogeneous computing "multi-arm coordination":

The first and most important thing is to improve the processing efficiency.

Heterogeneous computing can give full play to the flexibility of CPU/GPU in general computing, respond to data processing needs in a timely manner, and match special capabilities such as FPGA/ASIC to give full play to the performance of coprocessors and allocate computing resources reasonably according to specific needs.

The benefits of doing so are obvious, finding a balance between processing speed and power consumption, achieving efficiency and power saving.

For example, in the smartphone AI chip "two strong" Qualcomm and Kirin, the SoC of the former includes GPU accelerating 3D, ISP processing photos, baseband chip processing communication, DSP accelerating vector calculation, etc. Kirin 980's latest heterogeneous computing architecture is also optimized for system-wide integration based on CPU, GPU, NPU, ISP and DDR, seeking stronger performance and endurance.

Another advantage of heterogeneous computing is cost.

At present, neural network algorithms and corresponding computing architectures emerge endlessly. If ASIC architecture is continuously updated every time, it will sink to users and enterprises, which will lead to high cost of use and replacement.

Therefore, the best solution is to integrate multiple computing architectures together, and everyone will work collectively, and the life cycle will be much longer, and it will have greater advantages in industrial landing.

In addition to the strong improvement of hardware performance and industrial application, heterogeneous computing has a deeper value, that is, when the level of a single domestic isomorphic chip temporarily lags behind the international level, it is very likely to become a historical opportunity for Chinese chip curve overtaking.

How: How is the structure different?

Since heterogeneous computing is extremely important both at the national strategic level and for personal applications, how should different types of chips be put together?

Reflected in hardware, there are two main modes of development: one is chip-level (SoC) heterogeneous computing, such as Intel's KabyLake-G platform, which is to heterogeneously integrate Intel processors with AMD Radeon RX Vega M GPUs. Kirin 970, launched by Huawei last year, integrates NPU customized for deep learning on the basis of CPU and GPU to perform high-density calculations such as reasoning.

The other is Intel's hyperheterogeneous computing. Through EMIB, Foveros and other packaging technologies, the performance-proven chiplets are assembled into a packaging module.

Last year, Intel announced LakeField, a heterogeneous motherboard chip that integrates Intel's 10nm IceLake CPU and 22nm Atom small core. The integration of high-load and low-load processing cores is much smaller in size than simple and crude board integration.

From the perspective of hardware solutions, heterogeneous computing is the arrangement and combination of various processing cores, which seems to be similar to the technical difficulty of building blocks. However, there are many tricks to building an ideal coprocessor.

As a prerequisite, it is necessary to understand the specific capabilities of each processor, and then make exclusive products based on performance, power consumption, price, performance, etc.

In general, heterogeneous computing will choose CPU, GPU, FPGA, ASIC to arrange and combine. What advantages do they have?

Stable multi-energy cheap bowl CPU, is to calculate a brick, which needs to move where, is all heterogeneous solutions can not be abandoned.

Then, choosing who to group cp with becomes the key to differentiation.

GPU can execute highly threaded multi-process concurrent tasks, which can help CPU in large-scale tasks requiring complex control. For example, powerful personal computers, GPU is indispensable.

FPGA Chinese name is called "field programmable gate array," as the name implies, is reprogrammable wiring resources, so, can be used to achieve some custom special hardware functions. Moreover, its computational efficiency is higher than that of the first two companions, which is very suitable for processing AI algorithms and soon becomes the right arm of the CPU.

There is also a strong but not very public player, that is,"special custom integrated circuit"ASIC. Its programming method is to build circuits directly on physical hardware (gate circuits). Since there is no need to fetch instructions and decode, each time unit can focus on data processing and transmission, so it is the highest performance among all coprocessors, but the power consumption is the lowest. However, due to the need for low-level hardware programming, its customization is also expensive and lengthy, belonging to the legendary existence of the Jianghu.

At present, there are three main branches of heterogeneous computing, namely CPU+GPU, which is suitable for most general purpose computing and is the combination lineup most used by heterogeneous computing at present.

CPU+FPGA, high price, mostly used by enterprise users (such as Huawei, Baidu, IBM, etc.) for deep learning acceleration;

CPU+ASIC, less application, suitable for some large market, clear return on investment, a certain development cycle of the field, such as consumer electronics.

With the iteration of technology, we are likely to see multi-chip collaborative scenarios such as CPU+GPU+FPGA in the future. For example, Huawei's Atlas platform, which has just been released, can dynamically orchestrate the topology between multiple GPUs/FPGAs to further improve the overall performance of the system.

It has to be said that heterogeneous computing is opening the door to this new world and is bringing plenty of imagination to supercomputing, with the entire computing industry ecosystem actively involved.

However, sinking heterogeneous computing into a large industrial system is not as easy as we think.

WHEN: The rise of heterogeneous computing, not just technology

We have already introduced the past and present life of heterogeneous computing. But if you ask when you will see the real impact of heterogeneous computing, the answer may not be surprising. The reason is also very simple. The rise of heterogeneous computing depends not only on technology, but more importantly on active preparation from the application side.

However, under the reputation of "light of computing," heterogeneous computing has a high threshold for procurement, deployment and use. This leads to a number of challenges in its application:

For example, in terms of cost, if large-scale procurement cannot be realized, the procurement cost of heterogeneous computing chips is very high. Smartphone manufacturers can also bargain with the advantage of scale, while general enterprise users and individual developers have a particularly high purchase price for small quantities, especially customized boards such as FPGAs and AISCs, which are still far away from large-scale applications.

In addition, heterogeneous computing chip lead time is also very long. As the brain of artificial intelligence, GPUs around the world have been in short supply, and Nvidia has imposed restrictions on the number of chips each company buys every day. For programmable chips such as FPGAs and ASICs, it often takes months for enterprises to design, order and deliver hardware architecture due to unestablished programming standards and customization time.

As a result, the number and products are fixed, which may cause a mismatch between computing resources and actual applications on the one hand, and may have to continue to increase budgets due to the launch of new GPU/FPGA architectures. As a result, the upgrading cost of enterprises remains high, and naturally there are doubts.

Even if all the above problems are solved, the first-in-command makes money quickly, the chip arrives smoothly, and the hardware is successfully deployed, it is very likely that another situation will occur, that is, the offline GPU/FPGA and the online service cannot be connected, resulting in waste of resources and data island problems.

Oh, I'm so angry. Can I not do it myself? It's good to directly use the heterogeneous computing of cloud service providers?

Sadly, there are many pits. Because GPU, FPGA these ultra-high performance devices after cloud virtualization, performance loss is very serious, there will be a corresponding decline. The hardware optimization capabilities and solutions of different vendors vary greatly, and how to choose the right platform has become a problem.

In this way, the emergence and sinking of heterogeneous computing is simply a "minesweeping" game. Perhaps the true value heterogeneous computing brings to the digital world will emerge only when powerful vendors remove these hidden obstacles one by one.

When Chinese chip companies directly challenge the established giants with heterogeneous computing, the variables and possibilities in the process of industrial transformation will be more exciting than the technology itself.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.