16nm – shallow diving (not implying that anything is used)

insightful video:


Maxwell CPU throttling is managed with taking load from A53 to A57

Pascal CPU throttling is managed with taking load from A57 to Denver2

The 16nm production process is more efficient than 20nm and furthermore from 28nm

We know this from X1 shifting to 20nm as the only Maxwell built on 20nm.

Now nVidia has 16nm production as all Pascal are 16nm.

The rumored downclock and added that it still maintains the same 1TF output, are showing a gap of about 30% of missing performance, just about the performance increase with the shift from 20nm to 16nm (bearing in mind that Tegra SoC work with coherent CPU-GPU workload).

Pascal Signature: mixed precision GPU, 3D RAM, NVlink


  • mixed precision GPU


— note to reader – do a quick read

  • 3D High Bandwidth Memory


— Apparently nVidia develops their own 3D HBM (Vertically stacked, High Bandwidth Memory), that in some scenarios it can be cheaper to have, most effectively when die is limited

  • NVlink


— Speed compared to commercial PCIe Gen.3 – effective 1.3 to 5 times faster.

— Interconnect offers CPU-to-GPU high speeds and effective GPU-GPU and multiple GPUs-CPU scalability

Discussion started at here by rusenikolovski

Share this post