3 ms·

The GPGPU community is full of a decade+ of bright-eyed bushy-tailed AMD fans who were painfully and traumatically pushed into the NVidia camp after believing A

The GPGPU community is full of a decade+ of bright-eyed bushy-tailed AMD fans who were painfully and traumatically pushed into the NVidia camp after believing AMD's promises that it was totally ready for GPGPU (if only people would write OpenCL!) but then discovering the horrible truth that this was very much not the case. AMD's OpenCL was dire. I wasted months into that flaming piece of trash -- twice! -- and really wish I hadn't. Both times I had to toss my code, sell my card, eat the spread, eat the ebay tax, and pay the green tax to buy a lower-specced nvidia card, but that all sucked a lot less than tossing my very own programmer months into a dumpster fire. Ironically, the very openness that was supposed to win converts from the green camp actually sent people in the opposite direction because you could run your OpenCL on an nvidia card and find out that actually the problem wasn't with your code but rather with AMD's drivers. Hopefully now that AMD has money they can afford the headcount to fix their bugs and some more headcount to make up for 15 years of lost time in the applications that matter. Hopefully. But I trusted them twice and got burned, so this time I need to see proof before I jump. I need to see blender renders, I need to see torch trainings, I need to see NAMD (or insert science code here) running on a compatibility layer. It is unfortunate that the actual compatibility situation seems to have hardly moved at all in 10 years (it was right around the corner in 2013 too, you see) but the fundamentals are different now so I hold out hope.

dragontamer 3y ago
Lemme guess. You tried OpenCL 2.0 and found out it was a dumpster fire on AMD? OpenCL1.2 worked fine for all the test code I made for myself. But OpenCL 2.0 was... horrid. Unworkable. -------- My next step personally is to give DirectX 12 computer shaders a serious try. A lot of people are also talking about Vulkan shaders, as apparently that works well on Linux too. ROCm on Linux is fine enough. But this weirdness of having a feature on ROCm Windows (when ROCm Windows isn't even public yet?) is well... weird. I guess its cool that AMD has worked with the Windows team to get everything squared up, but it'd be nice if ROCm on Windows were available to the masses.

dragontamer 3y ago

Lemme guess. You tried OpenCL 2.0 and found out it was a dumpster fire on AMD? OpenCL1.2 worked fine for all the test code I made for myself. But OpenCL 2.0 was... horrid. Unworkable. -------- My next step personally is to give DirectX 12 computer shaders a serious try. A lot of people are also talking about Vulkan shaders, as apparently that works well on Linux too. ROCm on Linux is fine enough. But this weirdness of having a feature on ROCm Windows (when ROCm Windows isn't even public yet?) is well... weird. I guess its cool that AMD has worked with the Windows team to get everything squared up, but it'd be nice if ROCm on Windows were available to the masses.