NVIDIA publishes tech information on Pascal
After launching the Tesla P100 HPC board, NVIDIA has published official information on its new Pascal GP100 GPU, which will serve as the company’s flagship Pascal graphics processing unit.
According to the published information the GPU consists of 60 SMX units with each one of them having 64 stream processors. This means that the GPU includes 3840 stream processors or CUDA cores, as NVIDIA calls them. Pascal also has more FP64 units than the previous Maxwell and Kepler architectures – each SMX unit has 32 FP64 blocks for a total of 1920 units. The chip has a FP32 to FP64 ratio of 2:1, which allows the chip to effectively solve tasks that depend on double-precision calculations. The GP100 also has 4 MB of L2 cache, which improves on Maxwell’s 3 MB of L2 cache in the line’s top models. On the outside the GP100 resembles AMD’s Fiji chips – NVIDIA’s GPU is surrounded by four memory chips that are a bit larger than HBM memory chips. The GPU communicates with the memory over a silicon layer that sits in the middle. Moreover the HBM2 memory in the GP100 GPU has five layers – four memory layers and one base.
Unfortunately desktop video cards on the GP100 will not arrive soon since NVIDIA will launch graphics cards on the cut-down GP104 GPU first.