4 ms·
I hate the Cuda monopoly in machine learning right now. I hope some of the mature libraries (TF, Pytorch) start officially supporting AMD GPUs. Nvidia is looti
by screye 8y ago
I hate the Cuda monopoly in machine learning right now. I hope some of the mature libraries (TF, Pytorch) start officially supporting AMD GPUs.
Nvidia is looting customers that want to use GPUs for Machine Learning on the cloud (like AWS, GCP)
- shmerl 8y agoThere are some efforts to improve the situation: https://www.phoronix.com/scan.php?page=news_item&px=Red-Hat-Plumbing-Compute-Stack https://www.phoronix.com/scan.php?page=news_item&px=Red-Hat-...
- nabla9 8y agoNvidia is cashing in their investments for software and integration. It's AMD's job to make machine learning work in their GPU's. If they don't believe in it and spend the necessary time and effort, nobody else will. Radeon Open Compute Platform (ROCm) has existed for years but apparently it's not good enough. ps. Tensorflow has ROCm backend support. https://hub.docker.com/r/rocm/tensorflow/ https://hub.docker.com/r/rocm/tensorflow/ but is MI25 competitive? https://www.amd.com/en/products/professional-graphics/instinct-mi25 https://www.amd.com/en/products/professional-graphics/instin...
- ll01 8y agoNo clue but if consumer vega is competitive http://blog.gpueater.com/en/2018/04/23/00011_tech_cifar10_bench_on_tf13/ http://blog.gpueater.com/en/2018/04/23/00011_tech_cifar10_be... MI25 has a chance of being
- ndesaulniers 8y ago> Tensorflow has ROCm backend support. Why is it a separate repo rather than being contributed upstream to TF?
- nabla9 8y agoAMD is planning to upstream their work and some is already there, but are still behind in versions they support.
- singhrac 8y agoPyTorch has AMD support in the main repo: https://github.com/pytorch/pytorch/tree/master/tools/amd_build https://github.com/pytorch/pytorch/tree/master/tools/amd_bui...
- esistgut 8y agoIsn't ROCm a viable alternative?
- singhrac 8y agoNot really, at least not yet. It doesn't exactly provide the same API capabilities as CUDA yet (there are some unsupported functions) and no one has realistic comparisons of benchmarks but the consensus is that there's some performance gap between similar chips from Nvidia and AMD. Both are being actively worked on, and I expect to see this gap shrink in the next few months to a year.
- pjmlp 8y agoKhronos is to blame by focusing too much on C, without convient tooling. CUDA had C, C++ and Fortran support since the early days, followed by the PTX bytecode format for any compiler vendor that wanted to support CUDA on their languages. It was necessary to loose the race for them to focus on C++ and come up with SPIR and SYSCL. And tools still seem not to be on par with what NVidia offers.
- pcwalton 8y agoKhronos is a standards group. It's their job to make standards, not tooling. Individual vendors that are part of Khronos can make tooling.
- pjmlp 8y agoThen they made a bad standard to start with. Also regarding tooling, that isn't quite true, as they have engaged with LunarG for Vulkan, because they have learned from lack of OpenGL adoption, that the large majority of games developers wouldn't even bother without a proper SDK.
- geezerjay 8y ago> Then they made a bad standard to start with. Nonsense. You'll be hard-pressed to find a platform that doesn't support C or a competent system design or even graphics design engineer who isn't confortable with C. If their goal is to establish standards, they build upon standards.
- pjmlp 8y agoDevelopers jumping on CUDA due to lack of C++ and Fortran support on OpenCL proves how good that decision was. Likewise for game consoles, Windows and now Apple graphics APIs. Defining standards does not mean they must provide a lowest denominator single implementation. They can be defined in an IDL kind of way, like e.g. WebIDL, abstractly e.g. Internet RFCs, or define mappings to all major languages in the GPGPU field, namely C, C++ and Fortran.