Nvidia’s New DX11 Wonder Driver will be ”337.xx” and Arrive Sometime in April
Remember the wonder driver from Nvidia that showed massive performance increases on the DX11 API? Well we have some more information on that front. Nvidia has revealed in its GTC conference that it will be releasing said driver sometime in April and its nomenclature will be "337.xx".
Nvidia Releasing the "337.xx" DX11 Wonder Driver in April - CPU Overhead Solved on the Same API
Nvidia revealed in a talk at GTC that the DX11 API update that was showcased on the previous GTC session will arrive in the form of a Geforce Driver, more specifically the Geforce 337.xx Driver. It will be released sometime in April and will result in very significant performance gains inside the DX11 API itself. The new driver takes a well known approach to solve CPU overhead without moving to a new API (DX12 for example). The approach is called 'Tiled Resources' and a weak variant of the same is also used in Mantle API. This will also grant Developers more control at the Assembly level.
The numbers revealed by Nvidia were very impressive, especially considering that Low Level access is not possible without a rewrite of the API, this is pretty impressive. Ofcourse, the benchmarks themselves came under criticism because of the fact that a very high end processor was used to test a supposedly "CPU-Overhead-Reducing" Driver. However the performance gain on that very same processor just by shifting to this new driver is quite decent.
Now the numbers that were released showed a decrease in CPU overhead by a factor of about 4.This should result in a performance increase of leaps and bounds in multi-gpu configurations. Of course the fact remains, that if these charts were rigged or manipulated by Nvidia in any way, we will find out soon enough. When the driver is released and third party benchmarks emerge, it will become clear whether Nvidia was true to it word or not.
In another slide, NVIDIA has shown the performance scaling between GPUs and CPUs showcasing that the Single precision performance in GeForce GPUs has increased tremendously over the past four years with GeForce GTX 780 Ti pumping almost 5.5 TFLOPs of performance compared to Ivy Bridge which barely cross the 500 GLOPs barrier. The AMD Kaveri APU manages to do 856 GFlops but its still a mighty length behind the graphics units. Similar, the total bandwidth on GeForce GPUs has increased to 336 GB/s as opposed to 60 GB/s on Ivy Bridge. Broadwell is supposed to feature Crystalwell eDRAM which will increase bandwidth to around 100 GB/s but nothing compared to GPUs.
— Usman Pirzada (@usmanpirzada) March 30, 2014