They also provide acceleration of open source libraries like GROMACS, used for molecular dynamics simulation.
Of course AMD is making their own system that works on AMD and nVidia and Intel GPUs. It'll probably get locked down again if AMD gets market dominance.
One thing that strikes me is how it evolves with new features. Not just higher level libraries, but also more fundamental, low level stuff, such as virtual memory, standard memory models, c++ libraries, new compilers, communication with other GPUs, launching dependent kernels, etc.
At their core, OpenCL and CUDA both enable running parallel computing algorithms on a GPU, but CUDA strikes me as much more advanced in terms of peripheral features.
Every few years, I think about writing a CUDA program (it never actually happens), and investigate how to do things, and it’s interesting how the old ways of doing things has been superseded by better ways.
None of this should be surprising. As I understand it, OpenCL has been put on life support by the industry in general for years now.
Developing fast... OpenCL is much harder to learn than CUDA. Take someone who did some programming classes, explain how CUDA works and they'll probably get somewhere. Do the same thing with OpenCL and they'll probably quit.
That's what I thought as well, so the title on the website ("NVIDIA GPUs Enable Simulation of a Living Cell") is not really truthful then.
Due to cuda, tools, SDKs etc Nvidia is providing.
I'm not aware of anything similar at any other GPU company
I would think that Oracle's loss of a similar lawsuit with Java would be related.
NVIDIA said that they would be fine with it in the past, and ROCm HIP is just a (bad) CUDA API clone.
Not working in this space, I'm curious why this is the case. Is there something inherently better about CUDA? Or is it that Nvidia's performance is somehow better for these tasks? Or maybe something else?
One of the other posts mentions 2014 as a turning point. At that time, GPGPU stuff was entering the (scientific) mainstream and NVidia was all over academia, convincing people to try it out in their research. They handed out demo accounts on boxes with beefy GPUs, and ran an extremely generous hardware grant proposal. There was tons of (free) training available: online CUDA MOOCs and in-person training sessions. The first-party tools were pretty decent. As a result, people built a lot of stuff using CUDA. Others, wanting to use those programs, basically had to buy NVidia. Lather, rinse, repeat.
This is in stark contrast to the other “accelerator” vendors. Around the same time, I looked into using Intel Xenon Phi and they were way less aggressive: “here are some benchmarks, here’s how it works, your institution has one or two somewhere if you want to contact them and try it out.” As for OpenCL…crickets. I don’t even remember any AMD events, and the very broad standard made it hard to figure out what would work/work well and you might end up needing to port it too!
I'm sure that NVidia's ecosystem building played a role (I remember events in 2009 and books before that), perhaps even a big role, but it wasn't the only factor. I paid a steep price in 2014 and 2016 for incorrectly assuming that it was.
First of all, software that advertised OpenCL or AMD compatibility often worked poorly or not at all in that mode. Adobe creative suite just rendered solid black outputs whenever acceleration was enabled and forums revealed that it had been that way for years and nobody cared to fix it. Blender supported OpenCL for a while, but it was slower than CPU rendering and for a sticky reason (nvidia did the work to support big kernels with heavy branching and AMD didn't). Ironically, OpenCL mode had decent performance but only if you used it on an nvidia card.
The situation was even worse in scientific codes, where "OpenCL version" typically meant "a half-finished blob of code that was abandoned before ever becoming functional, let alone anywhere near feature-parity."
I quickly learned why this was the case: the OpenCL tooling and drivers weren't just a little behind their CUDA counterparts in terms of features, they were almost unusably bad. For instance, the OpenCL drivers didn't do memory (or other context object?) cleanup, so if your program was less than perfect you would be headed for a hard crash every few runs. Debugging never worked -- hard crashes all around. Basic examples didn't compile, documentation was scattered, and at the end of the day, it was also leagues behind CUDA in terms of features.
After months of putting up with this, I finally bit the bullet, sold my AMD card, bought an NVidia card, ate the spread, the shipping, the eBay fees, and the green tax itself. It hurt, but it meant I was able to start shipping code.
I'm a stubborn bastard so I didn't learn my lesson and repeated this process two years later on the next generation of cards. The second time, the lesson stuck.
It might be different now, but at that time, many of the users were not computer scientists, they were scientists with computers. Having an easier to use programming model and decent debugging tools means publishing more results, more quickly.
But it seems ROCm is getting better and it has tensorflow and pytorch support, so there's reasons to be hopeful to see some competition here.