Having a dual- or quad-core CPU isn’t very unique nowadays and CPUs with 12 and even 16 cores aren’t that uncommon. The Andromeda from Cerebras is a supercomputer with 13.5 million cores. The corporate claims it is among the greatest AI supercomputers ever constructed (however no longer the biggest) and will carry out 120 Petaflops of “dense compute.”
We aren’t positive concerning the technique, however additionally they declare multiple exaflop of “AI computing.” The pc has a material backplane that may care for 96.8 terabits according to 2nd between nodes. In keeping with a put up on Excessive Tech, the core era is a 3-plane wafer processor, WSE-2. One aircraft is for communications, one holds 40 GB of static RAM, and the mathematics aircraft has 850,000 impartial cores and three.4 million floating level gadgets.
The information is distributed to the cores and picked up by means of a financial institution of 64-core AMD EPYC 3 processors. Andromeda is optimized to care for sparse matrix computations. The corporate claims that the efficiency scales “virtually linearly.” This is, as you double the selection of cores used, you more or less part the full run time.
The system is to be had for far off use and price about $35 million to construct. Because it makes use of 500 kW at top run instances, it isn’t unfastened to function, both. Excessive Tech notes that the Frontier laptop at Oak Ridge Nationwide Labs is each greater and extra actual, however it price $600 million, so that you’d be expecting it to be extra succesful.
Maximum homebrew “supercomputers” we see are extra for studying tips on how to paintings with clusters than looking to hit this kind of efficiency. After all, when you’ve got a contemporary graphics card, OpenCL and CUDA will assist you to perform a little of this, too, however at a far lesser scale.