Note, the maximum allocation size is the same for both our OpenCL and Level Zero drivers. The reason why >4GB allocations work through the DPC++ runtime is because it explicitly relaxes the allocation limits.
Our current client GPUs report a 4GB limit because it lets us use 32-bit addressing arithmetic more aggressively, which improves performance. It's not something that's required by the spec.
This matches my understanding as well. A 16GB card would need to support at least 1GB allocations, and could report up to 16GB (no upper limit). Only less-than-4GB cards could have a less-than-1GB allocation limit.
I've filed an internal issue regarding the incorrect queries. Should be an easy fix. Feel free to file an issue on the compute-runtime GitHub also if you'd like. Thanks for raising these issues!
Following up: Linux should be much better with the latest drivers. I see 15+GB reported on my 16GB GPU. Windows still needs some work though and I don't have an ETA at this time.
I thought we fixed the amount of VRAM we report. I'll check and see. If you can send me the driver version you are using and whether you are on Windows or Linux that would help - thanks!
If you need help debugging and tracing your command buffer applications, the OpenCL Intercept Layer has full support for both the base command buffer extension and the mutable dispatch extension. github.com/intel/opencl-inte…#OpenCL
I also recently added command buffer emulation layer support for out-of-order command buffers for OpenCL implementations that support out-of-order queues. The same repo has sample applications that demonstrate how to use both new features. #OpenCL
Want to experiment with the new command buffer mutable dispatch extension right away? My command buffer emulation layer now supports this extension and should work with most OpenCL implementations. github.com/bashbaug/SimpleOp…#OpenCL
Epic road trip complete! 5000 miles, 34 nights (32 in a tent), 9 national parks, one flat tire 😥, countless memories. Pic from yesterday at Crater Lake.
Thanks for the Arc shirt @IntelGraphics! Got here just in time for sabbatical and I wore it today on our hike today in Great Basin National Park. #intelarc#IamIntel#grba
ALT View of Wheeler Peak from the Mather overlook in Great Basin National Park
I'm not going to stop working on the OpenCL Intercept Layer - there's simply too much well-tested functionality there that's not worth rewriting - but it's great to have layers as a first-class tool in our OpenCL toolbox!