問題文
Which characteristic of a workload predicts that moving it to accelerated hardware will pay off?
選択肢
- The workload spends most of its time waiting on network requests, and the device can issue those requests in parallel and keep many of them outstanding at once.
- The workload contains many branches whose outcome depends on the value of the previous element, because handling long dependency chains efficiently is exactly what a large number of parallel execution units is designed for and the hardware resolves the dependencies automatically.
- The same arithmetic is applied independently to a very large number of elements, so thousands of them can be processed at the same time.
- The workload is small enough to fit entirely in the device's memory, regardless of how much arithmetic it performs; residency in device memory is what removes the transfer step from the critical path.