30 ms·

Indeed, this is one way to think about it. However, PTX is an instruction set for a virtual machine, not the actual hardware. The true, hardware-specific assemb

Indeed, this is one way to think about it. However, PTX is an instruction set for a virtual machine, not the actual hardware. The true, hardware-specific assembly is called SASS (Streaming Assembly) and the PTX code is translated into SASS by the GPU driver (using ptxas) in a final compilation step. Unlike SASS, PTX is (mostly) forward compatible. I don't know the details about your CUDA 12 vs. 13 issue but I think it is not about hardware compatibility but more about the software stack. An application linked against CUDA 12 libraries and might not work with CUDA 13 libraries.

neuroelectron 1y ago
That's not much different than a modern CPU with an OS on top; where you have the OS doing some of the scheduling then the CPU is splitting up the instructions into microinstructions and then scheduling them again in finer detail (hyperthreading and such). Seems to me there must be a C-level syntax and compiler so you're not manually splitting up individual adds and such and is still capable of optimizing the math effectively. But if that were true, we wouldn't have AAA game studios going to NVidia to optimize their game engines for each individual game.