Having a twin or quad-core processor is not very unique today, and 12 and even 16-core processors aren’t that unusual. Cerebras’ Andromeda is a supercomputer with 13.5 million cores. The corporate claims it is one of many largest AI supercomputers ever constructed (though not the biggest) and might carry out 120 petaflops of “dense computing”.

We’re unsure of the methodology, however in addition they declare greater than an exaflop of “AI computing”. The pc has a material backplane able to dealing with 96.8 terabits per second between nodes. In keeping with an article on Excessive Tech, the core expertise is a 3-plane wafer processor, WSE-2. One airplane is for communications, one incorporates 40 GB of static RAM, and the maths airplane has 850,000 impartial cores and three.4 million floating level models.

Information is shipped to the cores and picked up by a financial institution of 64-core AMD EPYC 3 processors. Andromeda is optimized to deal with sparse matrix calculations. The corporate says efficiency scales “nearly linearly.” In different phrases, if you double the variety of cores used, you scale back about half of the entire execution time.

The machine is accessible for distant use and prices round $35 million to construct. Because it makes use of 500 kW at peak instances, it’s also not free to run. Excessive Tech notes that Oak Ridge Nationwide Labs’ Frontier pc is each bigger and extra correct, however value $600 million, so that you’d count on it to carry out higher.

A lot of the homebrew “supercomputers” we see are extra for studying find out how to work with clusters than attempting to realize that form of efficiency. In fact, you probably have a contemporary graphics card, OpenCL and CUDA will even permit you to do that, however on a a lot smaller scale.

Supply : https://information.google.com/__i/rss/rd/articles/CBMiHWh0dHBzOi8vaGFja2FkYXkuY29tL3RhZy9ocGMv0gEA?oc=5

Leave A Reply