Nvidia RTX 40-series to deliver 'up to 2–3 times increase in ray tracing' performance

pcgamer.com:

Nvidia has announced a massive bump in ray tracing performance with its next-gen Ada Lovelace graphics cards, and it's not all derived from the huge number of CUDA Cores that Nvidia has stuffed into its next-gen GPU (though it will have up to 18,000 of those).

One important part of that speed-up is a new technology that Nvidia's CEO Jensen Huang calls Shader Execution Reordering (SER). This «reschedules work on the fly» giving a 2/3X speed up for ray tracing on Ada Lovelace cards. Huang likens this to an engineering development of the same import as out-of-order execution was for CPUs, which is an important feature in pretty much every CPU today.

Then you have the obvious improvement: a new RT Core. The RT Core is the main driver of ray tracing performance in Nvidia's RTX graphics cards, and it had already seen one major overhaul with the RTX 30-series. With Ada Lovelace it will be bigger and better again, offering 200 RT TFLOPS performance and two times the ray-triangle intersection throughput. 

That's partially down to two new hardware units in the RT Core: a new opacity micromap engine, which «speeds up ray tracing of alpha test geometry by a factor of two times», and a new micromesh engine, which «increases geometric richness without the BVH build and storage cost.»

Then there's the new Tensor Core. Tensor Cores are the accelerator behind instructions used for machine learning and those similar, and this new 4th Gen Tensor Core offers 1,400 TFLOPS of Tensor programming and the Hopper FP8 transformer engine, right out of the Hopper architecture Nvidia previously announced for data centres.

So this whole change essentially amounts to a big shift in ray tracing performance, and Huang explains it has been necessary to