- English
-
EnglishDeutschItaliaFrançais한국의русскийSvenskaNederlandespañolPortuguêspolskiSuomiGaeilgeSlovenskáSlovenijaČeštinaMelayuMagyarországHrvatskaDanskromânescIndonesiaΕλλάδαБългарски езикGalegolietuviųMaoriRepublika e ShqipërisëالعربيةአማርኛAzərbaycanEesti VabariikEuskeraБеларусьLëtzebuergeschAyitiAfrikaansBosnaíslenskaCambodiaမြန်မာМонголулсМакедонскиmalaɡasʲພາສາລາວKurdîსაქართველოIsiXhosaفارسیisiZuluPilipinoසිංහලTürk diliTiếng ViệtहिंदीТоҷикӣاردوภาษาไทยO'zbekKongeriketবাংলা ভাষারChicheŵaSamoa日本語SesothoCрпскиKiswahiliУкраїнаनेपालीעִבְרִיתپښتوКыргыз тилиҚазақшаCatalàCorsaLatviešuHausaગુજરાતીಕನ್ನಡkannaḍaमराठी
E-mail:Info@YIC-Electronics.com
See the largest chip: 1.2 trillion transistors, 400,000 cores
On December 7, according to foreign media reports, Cerebras, based in Los Altos, California, has officially released a supercomputer named CS-1. Its core is a 16nm process technology and a wafer-sized processor. Array. The company named the 21.5 cm by 21.5 cm silicon board as WSE, which claims to be "the world's largest" chip. WSE pioneered the use of wafer-level integration, the number of transistors reached 1.2 trillion, and 400,000 cores.
It is reported that a high-end modern computer chip may have billions of transistors. But WSE has more than 1 trillion transistors. In addition, the 400,000 cores loaded by WSE can transfer 9000 trillion bytes of data per second; in contrast, the current Intel i9-9900k chip configured in personal computers has only 8 cores, which can transfer 40 billion per second. Bytes of data.
The supercomputer CS-1 is also very small. Currently the world's fastest supercomputing Summit has about 2.4 million cores. But Summit uses a traditional structure, using a heavily loaded circuit board, which weighs more than 340 tons and covers an area of 520 square meters. The CS-1 weighs only 50 kg, which is about the size of a domestic refrigerator. In terms of energy consumption, the Summit supercomputer consumes 1,000 times the power of the CS-1. It also consumes only 15-20 kilowatts of electricity. The peak requires 1000 times more energy.
The CS-1 supercomputer is mainly used for artificial intelligence operations such as machine learning. Its compiler is optimized to make data transfer between cores as efficient as possible. The CS-1 compiler improves transmission efficiency by matching the generated code structure with the hardware structure. In addition, because the core locations on WSE are only a few millimeters away from the memory they use, data flow from one part of the board to another is much faster than usual.
This wafer-level integrated semiconductor device itself is manufactured by TSMC. TSMC's manufacturing process is very precise, and each semiconductor device has only 150 to 200 defects. These are easy to ignore given the number of other transistors available. Of course, there are many other challenges for wafer-level integrated devices, such as data synchronization, power supply, heat dissipation, and data transmission effectiveness. However, if CS-1 can be successfully commercialized, wafer-level integrated semiconductor devices will eventually prove themselves.
It is reported that a high-end modern computer chip may have billions of transistors. But WSE has more than 1 trillion transistors. In addition, the 400,000 cores loaded by WSE can transfer 9000 trillion bytes of data per second; in contrast, the current Intel i9-9900k chip configured in personal computers has only 8 cores, which can transfer 40 billion per second. Bytes of data.
The supercomputer CS-1 is also very small. Currently the world's fastest supercomputing Summit has about 2.4 million cores. But Summit uses a traditional structure, using a heavily loaded circuit board, which weighs more than 340 tons and covers an area of 520 square meters. The CS-1 weighs only 50 kg, which is about the size of a domestic refrigerator. In terms of energy consumption, the Summit supercomputer consumes 1,000 times the power of the CS-1. It also consumes only 15-20 kilowatts of electricity. The peak requires 1000 times more energy.
The CS-1 supercomputer is mainly used for artificial intelligence operations such as machine learning. Its compiler is optimized to make data transfer between cores as efficient as possible. The CS-1 compiler improves transmission efficiency by matching the generated code structure with the hardware structure. In addition, because the core locations on WSE are only a few millimeters away from the memory they use, data flow from one part of the board to another is much faster than usual.
This wafer-level integrated semiconductor device itself is manufactured by TSMC. TSMC's manufacturing process is very precise, and each semiconductor device has only 150 to 200 defects. These are easy to ignore given the number of other transistors available. Of course, there are many other challenges for wafer-level integrated devices, such as data synchronization, power supply, heat dissipation, and data transmission effectiveness. However, if CS-1 can be successfully commercialized, wafer-level integrated semiconductor devices will eventually prove themselves.