Simulation by the world's largest chip 'Wafer Scale Engine' boasting 1.2 trillion transistors is explosive enough to exceed the laws of physics



The verification result of the simulation speed by the data center module ' Cerebras CS-1 ' equipped with the world's largest chip ' Wafer Scale Engine ' of 20 cm x 22 cm has been announced. According to the research team's announcement, 'Finally, it has become possible to make predictions based on simulations at speeds that exceed the actual laws of physics.'

[2010.03660] Fast Stencil-Code Computation on a Wafer-Scale Processor

https://arxiv.org/abs/2010.03660

Wafer scale Cerebras CS-1 flexes its muscles in scientific tasks --CPU --News --HEXUS.net
https://hexus.net/tech/news/cpu/146986-wafer-scale-cerebras-cs-1-flexes-muscles-scientific-tasks/


Trillion-transistor chip breaks speed record
https://techxplore.com/news/2020-11-trillion-transistor-chip.html


In 2019, startup company Cerebras Systems has partnered with the National Institute of Energy Technology in the United States to develop the giant chip 'Wafer Scale Engine'. This Wafer Scale Engine uses TSMC's 16nm process node, is equipped with 1.2 trillion transistors, and is talked about as the largest computer chip in history with approximately 400,000 sparse linear algebra computing cores. I did.

The largest computer chip in history with 1.2 trillion transistors will be developed --GIGAZINE



And the deep learning system that is about the size of a refrigerator equipped with this Wafer Scale Engine is 'Cerebras CS-1'.



The 18GB of SRAM in the core of the Wafer Scale Engine has a memory bandwidth of 9.6P (peta) B / s and is connected to a network fabric called 'Swarm' that operates at 100Pbps. The Cerebras System claims that the CS-1 is the most powerful AI computing system in the world.



Then, when a joint research team of Cerebras Systems and Energy Technology Research Institute conducted a combustion process simulation of the power plant with CS-1, it was an HP supercomputer 'Joule 2.0' equipped with an Intel Xeon chip with a total of 16,000 cores. The CS-1 reported that it was able to calculate at just 28 microseconds, which is more than 200 times faster than the calculation that took 6 milliseconds. 'No matter how many CPUs or GPUs you have, no supercomputer at the moment can match the performance of the CS-1,' the research team boasted.

As to why it was able to calculate 200 times faster than Joule, the research team said, 'The cache of Intel Xeon used in Joule is less effective in deriving performance from SRAM, and Intel Xeon is compared to the core of Wafer Scale Engine. Maybe it's because we can only get 40% performance. '

Cerebras Systems said, 'Working with CS-1 opens the door to major advances in scientific computing performance. CS-1 is powerful enough to simulate over 1 million fluid cells faster than in real time. For the first time, CS-1 can be used to simulate in real time, for example, in a thermal power plant based on data on current operating conditions. That is, CS-1 has the same physical laws. You can predict what will happen faster than you can produce results. '

In addition, Cerebras Systems has started development of CS-2 equipped with the next generation Wafer Scale Engine. According to the IT news site Hexus, by adopting TSMC's 7nm process node, the number of transistors is 2.6 trillion and the number of AI cores is 850,000, which is more than double that of CS-1.

in Hardware, Posted by log1i_yk