NVIDIA Appears To Finally Be Prepping OpenCL 2.0 Driver Support

[phoronix]

Phoronix: NVIDIA Appears To Finally Be Prepping OpenCL 2.0 Driver Support

The OpenCL 2.0 specification is going on four years old and it appears NVIDIA's proprietary drivers are finally getting ready to support this newer GPGPU computing specification...

http://www.phoronix.com/scan.php?page=news_item&px=OpenCL-2.0-NVIDIA-Preps

[tildearrow]

Slight miscapitalization: VUlkan

[Tabris]

OpenCL 2.0 symbols started appearing in the 378.xx series Linux binary shared library as well. Nothing in the json though yet.

[arteast]

Actually this driver version doesn't include OpenCL 2.0 support as the article claims; that would be too good for us. The driver still reports OpenCL 1.2 support only. The release notes say "New features in OpenCL 2.0 are available in the driver for evaluation purposes only.", i.e. they added preliminary support for few features from OpenCL 2.0 to their OpenCL 1.2 implementation; namely, Device side enqueue and Shared virtual memory. Also still no OpenCL-CUDA interop and no OpenCL-NVENC interop

[gururise]

Nvidia purposely crippled their OpenCL drivers and left them at 1.2 for years so that their CUDA could gain traction in the ML and Scientific community.. Now that Nvidia has a virtual monopoly in the scientific space, they seem to feel comfortable entertaining the idea of allowing OpenCL 2.0 into their drivers.

[birdie]

CUDA was miles better than the features offered by OpenCL so given the breadth of CUDA it made little sense for NVIDIA to support OpenCL.

They are a for profit company, they are not a pro Stallman idealistic company. They need to feed themselves and also their investors and shareholders.

Perhaps one day you'll realize that and stop throwing shit and groundless accusations at them.

Besides when something runs everywhere (AMD/Intel) there's no competitive advantage for them, vs. vendor locked CUDA which runs only on their GPUs.

[Tabris]

These symbols are new in libOpenCL.so.1.0.0 from 378.13:

OPENCL_2.0@OPENCL_2.0
clCreateCommandQueueWithProperties@OPENCL_2.0
clCreatePipe@OPENCL_2.0
clCreateSamplerWithProperties@OPENCL_2.0
clEnqueueSVMFree@OPENCL_2.0
clEnqueueSVMMap@OPENCL_2.0
clEnqueueSVMMemFill@OPENCL_2.0
clEnqueueSVMMemcpy@OPENCL_2.0
clEnqueueSVMUnmap@OPENCL_2.0
clGetPipeInfo@OPENCL_2.0
clSVMAlloc@OPENCL_2.0
clSVMFree@OPENCL_2.0
clSetKernelArgSVMPointer@OPENCL_2.0
clSetKernelExecInfo@OPENCL_2.0

[bug77]

Actually this driver version doesn't include OpenCL 2.0 support as the article claims; that would be too good for us. The driver still reports OpenCL 1.2 support only. The release notes say "New features in OpenCL 2.0 are available in the driver for evaluation purposes only.", i.e. they added preliminary support for few features from OpenCL 2.0 to their OpenCL 1.2 implementation; namely, Device side enqueue and Shared virtual memory. Also still no OpenCL-CUDA interop and no OpenCL-NVENC interop

"NVIDIA Appears To Finally Be Prepping OpenCL 2.0 Driver Support" too subtle for you?

Nvidia purposely crippled their OpenCL drivers and left them at 1.2 for years so that their CUDA could gain traction in the ML and Scientific community.. Now that Nvidia has a virtual monopoly in the scientific space, they seem to feel comfortable entertaining the idea of allowing OpenCL 2.0 into their drivers.

You must have missed some recent benchmarks right here on Phoronix where Nvidia's OpenCL 1.2 beats AMD's OpenCL 2.0. Other than that, you're spot-on.

[nevion]

They'll probably cave when they see AMD's platform is 100% less trouble to deal with and start competing back with OpenCL... my estimates are that this will take a year at least. Finally poking at the beast with a fork and knife.

[nevion]

Number one complaint on CUDA (from developers) besides the whole proprietary thing is it's super freaking slow to compile and extremely limited in some ways. About the only asset is it is/was easier to get started with. Hip will bridge that gap but OpenCL ftw. Such a better capability...