MIT develops CPU "Swarm" with chip dedicated to parallel processing of 64 cores faster and more efficient than before

A CPU equipped with a dedicated hardware chip for parallel programs that allows researchers at MIT to calculate a multi-core (many core) CPU 18 times faster than beforeSwarm"Was developed.

Parallel programming made easy | MIT News

MIT's Swarm chip architecture boosts multi-core CPUs

MIT creates chip that makes parallel programming faster, easier

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory announced a CPU with a dedicated chip that can efficiently write parallel programs called "Swarm". A major advantage of Swarm is that it can perform parallel processing extremely efficiently even though it has many CPU cores.

In order to improve the computation speed of the CPU, a method to raise the number of clocks has been used, but the clock number improvement has almost reached its limit about 10 years ago, and computation is performed in parallel by a plurality of cores "Multicore" has been promoted. However, when a plurality of CPU cores perform arithmetic processing in parallel, if it is necessary to access the memory data being processed by another core, even if the task processing is completed, the task of another core It is necessary to wait for the process, which is a useless standby time "overheadThere was a disadvantage that the processing speed greatly decreased. As the number of cores increases, the adverse effect of overhead accompanying this multicore expansion increases, and as the parallel program becomes more complicated, it is preferable to increase the number of cores to increase the calculation speed, The simple diagram was not established.

Swarm improves efficiency of parallel processing by multi-core by installing a dedicated hardware chip that determines the priority order of tasks. Since tasks can be timestamped, they are processed from tasks with high priority. As a result, if the task that has finished one core processing has a time stamp earlier than the task that has been processed by another core,conflictThe result can be written in memory immediately without causing it. Also, a high priority task (large task) has a mechanism to improve the efficiency of parallel processing by creating small tasks with low priority within their tasks.

Since Swarm has a dedicated chip for processing task priority, the amount of work to write parallel program seems to be extremely small, and compared with parallel program in conventional multi-core CPU without dedicated chip , It also has the merit of dramatically reducing the workload of programming.

Researchers at MIT made prototypes of the 64-core Swarm and performed arithmetic tests using six algorithms, it was confirmed that it was up to 18 times faster than the conventional multi-core CPU . Also, I showed the result that the amount of parallel program code that was necessary at that time was only 1 / 10th. Associate Professor Daniel Sanchez, who led Swarm development, said, "I suspect that development of multicore CPU does not progress as expected, because it was difficult to efficiently write parallel processing programs" Swarm says that this problem can be solved and there is a possibility of opening the way for development of a general purpose CPU having an enormous number of cores.

in Hardware, Posted by darkhorse_log