Hacker News Reader

Comment by jangid | Hacker News Reader

AMD lags not because of hardware only. There is very weak support for drivers and libraries in software. Intel and Nvidia spend huge amount of money in supporting library maintainers and thereby creating a lobby.

For example, look into the list of supported GPUs in the Github repositories of popular machine-learning libraries PyCharm, Tensorflow etc. AMD and OpenCL is nowhere as compared to Nvidia's CUDA.

View on HN

ROCm has been upstreamed to TensorFlow 2.0 now and it is gaining traction on AI platforms:

https://www.logicalclocks.com/blog/welcoming-amd-rocm-to-hop...

Also, the performance of AMD GPUs for deep learning is improving (50% in last 12 months through software alone). The Radeon VII ($600) is about the same performance as the Nvidia 2080Ti ($1100) - and the Radeon VII can be used in a data center (Nvidia force you to use the Volta GPUs - $9000 for a V100):

https://cdn.oreillystatic.com/en/assets/1/event/299/ROCm%20a...

Has this resulted in any increase in uptake of their GPUs? I'm usually the one to praise The Economist for covering technical subjects without embarrassing missteps but this:

Its gpus—which provide 3d graphics for video games and, increasingly, the computational grunt for trendy machine-learning algorithms—go up against those from Nvidia, whose revenues last year of $11.7bn were nearly twice those of amd.

seems wrong to me. They've been doing perfectly well in HPC tasks in supercomputers. And their CPUs have been doing great. But as far as I'm aware most machine learning work that moves away from NVidia moves to custom silicon rather than AMD GPUs. For the sake of openness I'd very much like AMD to start becoming a more popular option though.

On the other hand, Nvidia's Linux drivers are blobs (that don't even support Wayland) while AMD and Intel actively contribute to Mesa, and treat their users well (or at least much better than nvidia).

There is one advantage of nvidia proprietary driver, you get support of new cards on day one, with amd/intel it takes few months to have drivers in good shape.

I'm pretty sure amdgpu has supported AMD's recent cards (from the past few years) before they hit the shelves. Same for intel. It's always fun seeing the commits adding support for chips that you can't buy yet. And I've bought some very bleeding edge hardware from both companies in the past few years, with good results OOTB on Linux.

I look forward to the day that someone pulls the rug from under Nvidia like AMD did to Intel.

I'm not sure if it will happen, or when, but damn would it be satisfying.

There have been numerous situations just this year where I've stumbled on problem from having an AMD card, whereas Nvidias with their "blobs" have worked just fine. Gaming, video editing, even ML, they all seem to work perfectly on Linux with Nvidia, and AMD always has some problems. So at the very least this is contested. Nvidia seems much more supporting of the open source OS.

You don't have a laptop with a nvidia optimus card, do you? It's a shitshow of never ending problems with nvidia, while every amd laptop solution works as intended instantly after os installation.

I have had exactly the opposite experience, issues with Nvidia cards / drivers that I couldn't fix. Over the last 5~10 years AMD support on Linux (and other open source OSes) have been much better. To the point that for some of my workloads AMD on Linux is faster than AMD on Windows.

It is not surprising, AMD OpenGL drivers for Windows are horribly slow.

My experience in research taught me that NVidia makes a hostage-like situation. CUDA being a standard, it locked us into the proprietary driver, and overly expensive GPUs (or struggling to source used Titans) for double-precision computation. AMD stack not supporting CUDA at the time, their more powerful cheap double-precision GPUs felt like a forbidden fruit.

To top this off, the proprietary driver gave us permanent issues with OS integration, testing and deployment, and being distributed out of band, it was really difficult to get every developer on the same version. Each version of the GPU compiler had different and conflicting quirks, increasing the chaos of which developer/researcher used which branch of our software.

I would not recommend tying any serious work to NVidia hardware to any team.

It's crazy how amd still hasn't fixed their long long term driver issues. It's been a weak point for years and still is bad.

I'm look at their new cards but holding off because people are saying the only way to avoid crashes and black screens is to install on a fresh windows install. Yeah, no thanks.

Tbh at this point, I consider it a bit of hostile meme remaining from ATI times.

Driver support/reliability isn't something one could quantify in any meaningful way, as both companies have had plenty of botched driver releases [0] over the decades of their existence.

[0] https://www.tomshardware.com/news/Nvidia-196.75-drivers-over...

For a random person looking for those libraries: They meant *pytorch. Pycharm is a python IDE used for any python project.

Looks that Nvidia should put more money on his drivers for Linux. It are a piece of crap.

This is true for their GPUs but is a non issue for their CPUs.

Reply on news.ycombinator.com