Technology from the 80s: who revives waferscale-processors
The idea of waferscale-processors made a lot of noise in the 80s of the last century, but it turned out to be forgotten for decades due to the complexity of the production. Today they are trying to develop technology again. We tell who works in this area.
/ photo timo vn CC BY
In the production of modern processors, the silicon substrate is sawn into individual crystals, from which the chip is then assembled. Technology waferscale (WSI) involves the creation of processors (both CPU and GPU) on a single semiconductor wafer - it becomes one big "chip". This approach increases system performance: components are placed closer, so data is transferred faster.
')
In the 1970s and 1980s, WSI technology was proposed to be used to create chips for supercomputers. However, it was decided to abandon this idea - there were too many defects in the production of chips.
Only recently, WSI was again spoken about as a promising direction. For example, it is expected that such devices will help improve the performance of IT infrastructure in data centers. Interfaces for transferring data between server components are often a bottleneck. Waferscale systems, by placing part of the components on a single substrate, can help solve the bandwidth problem.
The organizations developing waterscale processors today are the University of California at Los Angeles and the University of Illinois at Urbana-Champaign. In early February of this year, university staff suggested creating a system that would combine dozens of graphics processors on a single silicon one. The project was named Silicon Interconnect Fabric, or Si-IF.
The engineers managed to create two prototypes for 24 and 40 GPUs. The developers conducted comparative tests with classical devices (with the same number of computational modules) - the performance of Si-IF devices was 2.5–5 times higher. Potentially, the technology can speed up the operation of servers based on graphics processors in data centers.
Since 2012, the Taiwanese manufacturer TSMC has been working on Waferscale processors. Their technology is called CoWoS - “chip on a plate on a substrate”. A silicon bridge and memory, FPGA and memory or network controllers are located on the silicon bridge (interposer). The bridge is then installed on a circuit board and “packaged” like a normal processor. The solution is already used in real devices, for example, Nvidia graphics chips from the Volta series.
Also, engineers are engaged in the development of technology in the framework of the WIPE project, which is part of the European initiative to create innovative solutions for Horizon 2020 . WIPE has seven participants, including IBM, the Belgian iMinds Center for Microelectronics, and a number of research institutes.
/ photo Enrique Jiménez CC BY-SA
The organization proposes to create a processor for working with fiber optic networks. In it, the module for converting an optical signal into an electrical one will be located above the silicon wafer with chips and connected to it directly through metal contacts. Researchers expect that this architecture will consume less power and run faster than traditional systems with separate transceivers and processors.
Experts expect that the development of WSI-processors will help data centers to increase the capacity of servers for resource-intensive tasks, such as training neural networks. According to some information, the startup Cerebras is already working on the waferscale processor for machine learning - although the company itself hides the details of the development.
Another advantage of WSI is its ease of management. For operating systems and other programs, waferscale devices do not look like several separate processors, but, for example, like one large GPU. This simplifies the work of programmers and system administrators who do not need to individually configure each module waferscale-system.
But waferscale processors have a limitation that makes it difficult to scale the IT infrastructure. Power systems in WSI chips require the installation of voltage regulators. These blocks take up space on which additional computational modules could be placed. Another limitation for waferscale-processors is associated with cooling devices. The high density of modules on the chip complicates heat dissipation.
But according to experts, these problems can be overcome and now is the right time for the revival of WSI. Since the 1980s, the production technology of processors have become more perfect, the number of defects has decreased. Therefore, we can expect that in the near future there will be more and more projects dedicated to the development of waferscale devices.
What we write about in the corporate blog:
And in the Telegram channel:
/ photo timo vn CC BY
What is waferscale
In the production of modern processors, the silicon substrate is sawn into individual crystals, from which the chip is then assembled. Technology waferscale (WSI) involves the creation of processors (both CPU and GPU) on a single semiconductor wafer - it becomes one big "chip". This approach increases system performance: components are placed closer, so data is transferred faster.
')
In the 1970s and 1980s, WSI technology was proposed to be used to create chips for supercomputers. However, it was decided to abandon this idea - there were too many defects in the production of chips.
For example, in this area, a startup Trilogy Systems worked, which then became the most funded in Silicon Valley - it received $ 230 million investment. But the company failed to form a reliable process technology. In 1985, Trilogy ceased research on WSI and went bankrupt, which adversely affected the reputation of waferscale processors in general.
Only recently, WSI was again spoken about as a promising direction. For example, it is expected that such devices will help improve the performance of IT infrastructure in data centers. Interfaces for transferring data between server components are often a bottleneck. Waferscale systems, by placing part of the components on a single substrate, can help solve the bandwidth problem.
Who develops technology
The organizations developing waterscale processors today are the University of California at Los Angeles and the University of Illinois at Urbana-Champaign. In early February of this year, university staff suggested creating a system that would combine dozens of graphics processors on a single silicon one. The project was named Silicon Interconnect Fabric, or Si-IF.
The engineers managed to create two prototypes for 24 and 40 GPUs. The developers conducted comparative tests with classical devices (with the same number of computational modules) - the performance of Si-IF devices was 2.5–5 times higher. Potentially, the technology can speed up the operation of servers based on graphics processors in data centers.
Since 2012, the Taiwanese manufacturer TSMC has been working on Waferscale processors. Their technology is called CoWoS - “chip on a plate on a substrate”. A silicon bridge and memory, FPGA and memory or network controllers are located on the silicon bridge (interposer). The bridge is then installed on a circuit board and “packaged” like a normal processor. The solution is already used in real devices, for example, Nvidia graphics chips from the Volta series.
Also, engineers are engaged in the development of technology in the framework of the WIPE project, which is part of the European initiative to create innovative solutions for Horizon 2020 . WIPE has seven participants, including IBM, the Belgian iMinds Center for Microelectronics, and a number of research institutes.
/ photo Enrique Jiménez CC BY-SA
The organization proposes to create a processor for working with fiber optic networks. In it, the module for converting an optical signal into an electrical one will be located above the silicon wafer with chips and connected to it directly through metal contacts. Researchers expect that this architecture will consume less power and run faster than traditional systems with separate transceivers and processors.
What are the prospects
Experts expect that the development of WSI-processors will help data centers to increase the capacity of servers for resource-intensive tasks, such as training neural networks. According to some information, the startup Cerebras is already working on the waferscale processor for machine learning - although the company itself hides the details of the development.
Another advantage of WSI is its ease of management. For operating systems and other programs, waferscale devices do not look like several separate processors, but, for example, like one large GPU. This simplifies the work of programmers and system administrators who do not need to individually configure each module waferscale-system.
But waferscale processors have a limitation that makes it difficult to scale the IT infrastructure. Power systems in WSI chips require the installation of voltage regulators. These blocks take up space on which additional computational modules could be placed. Another limitation for waferscale-processors is associated with cooling devices. The high density of modules on the chip complicates heat dissipation.
But according to experts, these problems can be overcome and now is the right time for the revival of WSI. Since the 1980s, the production technology of processors have become more perfect, the number of defects has decreased. Therefore, we can expect that in the near future there will be more and more projects dedicated to the development of waferscale devices.
What we write about in the corporate blog:
- How IaaS helps grow your business: three tasks that the cloud will solve
- What is behind the term vCloud Director - an inside view
- Distributed firewall in vCloud Director 8.20: solution features
And in the Telegram channel:
Source: https://habr.com/ru/post/441996/
All Articles