There are many types computing, calculating edge, heterogeneous computing, quantum computing, and more numerous. Different computing have different focus areas and specific uses, their development is also a great difference, but overall, the main method of modern computing is the future of the world and solve problems.
What is the significance of heterogeneous computing?
Speaking of heterogeneous computing, in fact, for people who understand the computer industry is not new. As computing to diversify, an increasing number of scenarios started to introduce a variety of different computing unit CPU, DSP, GPU, ASIC, FPGA and the like to speed up the computation, whereby, heterogeneous computing emerged. The core point is that heterogeneous computing "heterogeneous" word, it means a different process architectures, different instruction sets, different hardware capabilities combine to solve the problem, which is heterogeneous computing.
Gave birth to a wide range of computing needs heterogeneous computing
Tips: Encyclopedia definition of heterogeneous computing: heterogeneous computing is a special form of parallel and distributed computing, which is either a stand-alone computer can support a single mode and simd mimd manner, or by a high-speed network with interconnected a group of independent computer to complete computing tasks. It can be used to coordinate the performance of different structures to different computing machines to meet demands, and the code (or code segments) can be performed to obtain the maximum overall performance mode.
Why use heterogeneous computing?
So why use different hardware architectures of the process, rather than the same hardware architecture process to solve the problem?
This one might actually have a lot of people there are some misconceptions, the semiconductor chip with the CPU equated. The semiconductor chip manufacturing process but in fact, not only the process comprising a processor, further comprising a storage, communications, graphics chip. And each chip is not entirely by the vendor design, production, packaging, various different technical strength, advancing chip manufacturing process technology in their respective fields of speed is different. Such as a processor chip into the 14nm process node, but may also GPU chip 22nm process node, the communication chip may still 28nm process node, and the different structure between the different chips, so if there is no heterogeneous technologies, it is difficult to these different specifications of the chip package to a motherboard used.
What types of heterogeneous computing? (Configuration of view mode)
Heterogeneous computing is not what emerging technologies, from the beginning of the 1980s, heterogeneous computing has emerged and began to be trying to put into practical applications. Past, heterogeneous computing is divided into two types:
One is chip-level (SoC) heterogeneous computing;
Another board-level integration of heterogeneous computing.
Both the name suggests, chip-level (SoC) heterogeneous computing is the different processes, different architectures of heterogeneous chip to solve computing problems. For example, last year Intel launched on Pluto Canyon KabyLake-G platform, Intel processor is to be heterogeneous with AMD Radeon RX Vega M GPU, computation and graphics computing to solve problems.
Heterogeneous computing board equally well understood that different isomers motherboard functions to solve the problem by calculating a high bandwidth connection.
What are the advantages of traditional heterogeneous computing and inadequate
However, conventional heterogeneous computing, there are some advantages and disadvantages between each.
Chip-level heterogeneous computing in power consumption and performance have a good advantage, but it requires the designer to have a very deep understanding of the application load while before heterogeneous SoC design considerations to spend at least 18 months to do a different structure chip in the end is not worth it. In addition, the chip-level heterogeneous once completed can not be changed, if the 18-month period customer needs change, so they have to spend quite a long period to re-design, making a heterogeneous chip, the overall labor costs, time cost is very high, and in terms of flexibility is obviously inadequate.
Compared heterogeneous chip-level, board-level heterogeneous computing advantage is more flexible than the former. But the disadvantages are also obvious, after all, are connected between the motherboard and use the motherboard, one is born with relatively large volume, bandwidth and power consumption addition is the connection between the board and the board are difficult to achieve the optimal solution.
So, you want to heterogeneous computing breakthrough, or is adapted to current and future application needs, we need new technologies to support.
Revisit background heterogeneous computing
In addition Some may ask: "to PC, for example, non-mainstream heterogeneous computing is at this stage, to consider why the fuss over heterogeneous computing it?"
In fact, here to give an account of the background simple semiconductor chip manufacturing process. As we all know, since nearly a year, the semiconductor industry is moving from 14nm process node to 10nm, evolved Node 7nm, some time ago but news leaked that Apple has in the chip layout 5nm. So I have not thought about evolution in the process to 5nm nodes, 3nm even after 1nm, the development of miniature technology also can meet such a rapid iterative evolution of node it? If you can not meet, then there is no other technology can make up for it?
In fact, this is the stage of industry revisit a large background of heterogeneous computing. That is when the process node evolution is slowing down, the new framework R & D costs increased, then the calculation to solve a larger scale, the higher the load, the isomerization is a very good and effective solutions.
All along, the development of the semiconductor chip industry are focused on cutting-edge technology and continuously push forward the process node aspects of the evolution of the system, but when approaching the physical limits of process technology, and a short time when technology can not break through, we should change the thinking to meet computing requirements, rather than blindly like a dead end, "rushed forward."
Therefore, in this context, Intel started to pay attention not only to re-heterogeneous computing, but also on the basis of its proposed super heterogeneous computing concept.
Ultra heterogeneous computing "super" Where?
Relative to the chip-level and board-level heterogeneous computing, super core heterogeneous computing point is high flexibility and customization. President of Intel China Research Institute Song Jijiang in the interpretation of super heterogeneous computing "super" word to say: " 'Super' will be over in a lot on the existing, proven quite different nodes good Chiplet integrated in one package at this level can guarantee that the volume is small, it can go any lower-power control of a number, then you can enjoy higher bandwidth and shorter delay. on certain costs much cheaper than the board set a combination, but both fast and flexible, and it may even be cheaper than SoC. If you do 10nm SoC chip heterogeneous, then the cost may not be cheap, but now some 10nm and 14nm, and even 22nm chip integrated use, so that you can be well controlled cost."
Thus, Intel's "super heterogeneous computing" concept, is to achieve by different calculation module packaging technology integration, by EMIB, Foveros these 2D, 3D packaging techniques plurality Chiplet (small chips) fitted into one package module on the one hand, unlike heterogeneous SoC technology so complicated, but also avoid the problem of lack of flexibility due to the long period; on the other hand is smaller than the volume of heterogeneous traditional board level.
Intel's first stacked Foveros 3D design of heterogeneous motherboard chip LakeField
If the concern, then the beginning of CES may recall, Intel also announced 10nm IceLake platform, also announced a new 3D packaging technology by Foveros build LakeField heterogeneous motherboard with integrated Intel 10nm IceLake CPU on this motherboard and small 22nm Atom core is small. The former is responsible for high load calculation processing, which is responsible for low load operation, the force can be considered reasonable distribution and consumption. And you can see, in this way heterogeneous board, at the same time with full PC functionality, but also to ensure that the small size of the design. Meanwhile integrated on a motherboard, you can avoid bandwidth limitations, to achieve higher efficiency.
LakeField heterogeneous Motherboard
The main problem to be solved by itself heterogeneous computing
After heterogeneous computing needs to solve many problems, but the most important is the interconnection between different chips, as well as the different chips together, how to solve the power consumption and heat dissipation. For now, Intel gives the answer with a reference value by Foveros 3D package to create LakeField.
Epilogue
Ultra heterogeneous computing proposed concept, largely for Intel process node evolution is slowing down, the gradual increase in semiconductor R & D costs is an effective coping strategies. Ultra heterogeneous manner due to the more flexible and can be customized features, it can integrate different processes of the chip, and therefore play a significant role in cost savings.
In practice, the past is over heterogeneous technology has been tested, proven, industry have had to adapt to the chip immediately use it, and therefore avoid the risk of the validation phase. For example, now 22nm, 28nm chip manufacturing process has been very mature, and by doing heterogeneous packaging and 14nm, 10nm chip, maturity and performance of the whole industry will not cause problems. This is a super heterogeneous computing value.
