At the 2016 GPU Technology Conference, nVidia announced a new top-class GPU – the Tesla P100. They dubbed it ‘the most advanced hyperscale datacenter GPU ever built’, and it’s release gives us an all new insight on the design and features of the new Pascal architecture which is to feature in the upcoming GTX 1080/1070 graphics chipsets.
The GP100 GPU is said to use HBM2, which provides memory throughput of 720GB/s peak bandwidth in it’s 16GB memory configuration. HBM2 offers quadruple the memory density of HBM1 and double the bandwidth per stack allowing the GP100 to effectively leverage all 3584 of it’s CUDA cores.
The Pascal SM architecture features a 1:2 Ratio of 64-bit Floating Point paths allowing double precision floating point performance at half the single point rate. This new streaming multiprocessor cores should also result in more efficient execution of code. It’s 3584 CUDA cores are clocked at 1328MHz Base, which gives the GP100 a total compute performance of at least 10.6TFLOPs and subsequently 5.3TFLOPs of double precision compute, all manufactured on a 16nm process node.