On demand paging

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Hi,

I know that Pascal supports on demand paging i.e. when declaring an array with unified memory features the host cpu can access the array when the kernel is running which was not possible on Maxwell and Kepler.
However I am not sure if this is feature works on TX2 or not because some pascal features like system wide atomics do not work on tx2.
Could anyone give an update soon if all GP100 unified memory features that simultaneously allow cpu and gpu to access the same pool of memory allocated using cudaMallocManaged is available on the TX2 before I order this board? Thanks

Best,
Trinayan

2

Hi,

We are checking internally. Will update to you later.
Thanks.

3

Hi,

Similar to system-wide atomic[op], concurrentManagedAccess only support discrete GPUs on Linux.
Thanks.

4

Hi,

Thanks. I guess this information is also not present in the cuda Manual for reference.Might be helpful to put this.
It is very strange and also sad that most of the advanced features presented in the Whitepaper that make the GTX 1080 great are absent in the TX2 which has the same underlying architecture. I hope we get these features in the future. Thanks.

Best,
Trinayan

5

Hi,

We have added this information to document when topic1009422.
The document will be updated once a new release is available.

These kinds of operation are system-related. Although we have enabled these for x86_64, for ARM, it requires entirely different handling.
That’s why concurrentManagedAccess doesn’t support TX2.

Thanks for the feedback.