The CUDA 7.0 RC webinar just finished and it was noted at the end that TK1 CUDA support won’t go beyond CUDA 6.5.
I believe this is because 32-bit Linux support has been dropped in CUDA 7.0.
I am guessing that the 64-bit Denver variant and obviously the forthcoming X1 will receive CUDA 7+ support.
Interesting… but
On the cuda 7.0 download page http://go.nvidianews.com/K00oLOF306jQ0AEAQ03M0N0 it says:
Linux FAQ
Q: Where can I find the CUDA 7 Toolkit for my Jetson TK1?
A: CUDA 7 Toolkits for Jetson TK1 will available as part of a new L4T release after CUDA 7 production release.
Perhaps there will be a CUDA 6.5.1 or 6.6…which would mean no API changes but bug fixes…we’ll have to wait and see.
This is a typo on the download page. Thank you for pointing this out.
The typo will get fixed soon on the FAQ. In the meantime, to target Jetson TK1, developers should continue to use CUDA 6.5, and there are no plans to support that product with CUDA 7.0.
This is not related to dropping 32-bit Linux, as that change was limited to x86-based systems.
So that kind of sucks seeing as how TensorFlow is now available but REQUIRES CUDA 7.0… Do you think anyone might be able to create a build?
The trick is that CUDA 7 requires a 64-bit environment, and Jetson TK1 is 32-bit. The Jetson TX1 has just been announced and is 64-bit, so this should directly support CUDA 7. With TensorFlow being open source it could be modified for CUDA 6.5…to know how much trouble this would be would likely require an actual conversion attempt by someone experienced in both CUDA 6.5 and 7. It’s hard to predict if there will be enough interest in a 32-bit version or need…it is possible though.