GTX 1070 Clocks Throttle Reasons Unknown - Active

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi all!

I’m using
Ubuntu 16.04LTS
GTX 1070
CUDA Toolik 8.0
Driver Version: 375.26

During computation GPU-Util never reach 100%(that can be explained by inefficient implementation) and nvidia-smi -q -d PERFORMANCE returns

==============NVSMI LOG==============

Timestamp : Tue Jan 10 23:10:38 2017
Driver Version : 375.26

Attached GPUs : 1
GPU 0000:01:00.0
Performance State : P2
Clocks Throttle Reasons
Idle : Not Active
Applications Clocks Setting : Not Active
SW Power Cap : Not Active
HW Slowdown : Not Active
Sync Boost : Not Active
Unknown : Active

#2

I am running the same bandwidth test with GTX 1070 on LTS 16.04 driver 375.26 but using CUDA Toolkit 7.5

I see a distinct drop to 140 GB/s when getting close to 8000MB memory use. Your bandwidth seemed to uniformly stay above 200 GB/s. Puzzling.

How exactly did you build the bandwidth test utility? (command line)