Google unveiled their Tensor Processing Unit custom x86 chips at last year’s Google I/O, and has now published a white paper going over all the details of how they work, as well as presenting workload-optimized benchmarks that show TPUs doing what they do best about 15 to 30 times faster than Google’s previous rigs, which used Intel’s Haswell-era processors in tandem with NVIDIA’s K80 GPU family. The specially made chips leave normal computing equipment in the dust for the sort of machine learning tasks that Google has them doing due to an insane amount of customization and optimization, but Google did not perform any normal synthetic benchmarks with them, which means that we still have no idea how powerful they actually are in relation to what’s in most desktop and laptop PCs.
In addition to being vastly better at machine learning and neural networking tasks than typical processors, TPUs get Google more bang for their buck when it comes time to pay the power bill. Depending on the application and what it’s being compared to, TPUs were shown to outstrip their contemporaries by up to 3 times over in operations per watt. Incremental performance saw the most dramatic difference, with a TPU putting out 196 teraoperations per watt, versus 83 from the closest competitor. With the raw speed on offer by TPUs, it’s not hard to imagine storage and RAM becoming bottlenecks, so we can likely expect even better numbers as those get faster and Google upgrades their data centers.
While Google does not plan to release TPUs to the general public, there are some definite consumer implications here, and big ones at that. The most obvious is that other companies are likely going to see the massive success Google is having with the tactic of building custom, task-optimized processors, and try that themselves, which could lead to vastly optimized, faster, and less power-hungry gadgets in the near future. Specialized gizmos like IoT devices, for example, could benefit quite a bit from custom development. Google even actively encouraged such development, telling chip designers to “set the bar higher” and surpass the TPU with their creations. So, while you won’t be seeing a TPU in your future superphone, server, or gaming PC any time soon if at all, specialized chips for each component or each job inside the system are a possibility, and could bring huge performance and power improvements.
The post Google Publishes An Analysis Of Their Custom TPU Chips appeared first on AndroidHeadlines.com |.
【Top 10 Malaysia & Singapore Most Beautiful Girls】Have you follow?
