The following chart is an updated version of the one published in our GeForce GTX 1660 review, which illustrates TU117’s massive improvement to half-precision throughput compared to GeForce GTX 1060 and its Pascal-based GP106 chip. TU106, TU104, and TU102 boast double-rate FP16 as well through their Tensor cores, so TU117’s configuration serves to maintain that standard through hardware put in place specifically for this GPU.
In TU117, Nvidia replaces Turing’s Tensor cores with 128 dedicated FP16 cores per SM, which allow GeForce GTX 1650 to process half-precision operations at 2x the rate of FP32. In between, it's free to issue a different instruction to any other unit, including the INT32 cores. Because Turing doubles up on schedulers, it only needs to issue an instruction to the CUDA cores every other clock cycle to keep them full. Four of those 16-core groupings comprise the SM, along with 96KB of cache that can be configured as 64KB L1/32KB shared memory or vice versa, and four texture units. The newer architecture assigns one scheduler to each set of 16 CUDA cores (2x Pascal), along with one dispatch unit per 16 CUDA cores (same as Pascal). Turing’s Streaming Multiprocessors are composed of fewer CUDA cores than Pascal’s, but the design compensates in part by spreading more SMs across each GPU.
0 kommentar(er)
