260GTX nvidia-drivers-190.42-r3 - powermizer issue

[gio___]

to mrn and cjg:
the laptop is a rebrand of a Clevo w860CU. in all my nvidia-bug-report you can find more specific information.

summarizing:
-CPU i7-820QM
-4Gb ram DDR3 1333Mhz
-nvidia GTX 260M

- Gentoo Linux, kernel 2.6.34
- Xorg-server 1.8.1-r1
- Awesome WM

Ok glxgears is not a bencmark but is quite indicative of the overall perfomance. In the results that I posted xcompmgr (the compositing manager) was disabled.

I have CUDA Sdk installed so tell me which test you prefer or what other benchmark program you want.

[cjb]

Originally Posted by gio___

Ok glxgears is not a bencmark but is quite indicative of the overall perfomance. In the results that I posted xcompmgr (the compositing manager) was disabled.

I have CUDA Sdk installed so tell me which test you prefer or what other benchmark program you want.

Thanks - I'm interested in the FDTD3d example myself; something like

./FDTD3d --dimx=144 --dimy=144 --dimz=144

for which I get:

gtx260m - 284 MPoints/s
gt240 - 754 MPoints/s
c1060 - 1909 MPoints/s

The gtx260m should really be close to the gt240. This one's not a sterling benchmark either, but it's fairly close...

[gio___]

Originally Posted by cjb

Thanks - I'm interested in the FDTD3d example myself; something like

./FDTD3d --dimx=144 --dimy=144 --dimz=144

for which I get:

gtx260m - 284 MPoints/s
gt240 - 754 MPoints/s
c1060 - 1909 MPoints/s

The gtx260m should really be close to the gt240. This one's not a sterling benchmark either, but it's fairly close...

./FDTD3d --dimx=144 --dimy=144 --dimz=144
./FDTD3d Starting...

Set-up, based upon target device GMEM size...
getTargetDeviceGlobalMemSize
cudaGetDeviceCount
cudaGetDeviceProperties
calloc host_output
malloc input
malloc coeff
generateRandomData

FDTD on 144 x 144 x 144 volume with symmetric filter radius 4 for 5 timesteps...

fdtdReference...
calloc intermediate
Host FDTD loop
t = 0
t = 1
t = 2
t = 3
t = 4

fdtdReference complete
calloc device_output
fdtdGPU...
cudaGetDeviceCount
cudaSetDevice (device 0)
cudaMalloc bufferOut
cudaMalloc bufferIn
set block size to 16x16
cudaMemcpy (HostToDevice) bufferIn
cudaMemcpy (HostToDevice) bufferOut
cudaMemcpyToSymbol (HostToDevice) stencil
cudaEventCreate
cudaEventCreate
GPU FDTD loop
t = 0 launch kernel
t = 1 launch kernel
t = 2 launch kernel
t = 3 launch kernel
t = 4 launch kernel

cudaThreadSynchronize
cudaMemcpy (DeviceToHost)
cudaEventElapsedTime

FDTD3d, Throughput = 179030.9718 MPoints/s, Time = 0.00002 s, Size = 2820096 Points, NumDevsUsed = 1, Blocksize = 256

[cjb]

Originally Posted by gio___

FDTD3d, Throughput = 179030.9718 MPoints/s, Time = 0.00002 s, Size = 2820096 Points, NumDevsUsed = 1, Blocksize = 256

...and it reported PASSED at the end? This number would imply a minimum memory bandwidth of 716 GB/s (the Tesla C1060 has a peak of 106, and the gtx60m has 61; good if we can sustain 10 or so) and the FDTD algo is not quite that efficient. Something's not quite right...

[gio___]

Originally Posted by cjb

...and it reported PASSED at the end? This number would imply a minimum memory bandwidth of 716 GB/s (the Tesla C1060 has a peak of 106, and the gtx60m has 61; good if we can sustain 10 or so) and the FDTD algo is not quite that efficient. Something's not quite right...

you're right:

CompareData (tolerance 0.000100)...
Data error at point (4,4,4) 0.047833 instead of 10.829188

FAILED


!!! Error # 0 at line 48 , in file src/FDTD3d.cpp !!!

Exiting...
-----------------------------------------------------------

[cjb]

Originally Posted by gio___

you're right:

CompareData (tolerance 0.000100)...
Data error at point (4,4,4) 0.047833 instead of 10.829188

FAILED


!!! Error # 0 at line 48 , in file src/FDTD3d.cpp !!!

Exiting...
-----------------------------------------------------------

Well, believe me, I'd be one rich little sucker tomorrow if I could get that number for real. Meantime, try changing 144 to 128 (the smallest setting it allows) for all three numbers, and see if you get a good result; thanks

[gio___]

Originally Posted by cjb

Well, believe me, I'd be one rich little sucker tomorrow if I could get that number for real. Meantime, try changing 144 to 128 (the smallest setting it allows) for all three numbers, and see if you get a good result; thanks

# ./FDTD3d --dimx=128 --dimy=128 --dimz=128
./FDTD3d Starting...

Set-up, based upon target device GMEM size...
getTargetDeviceGlobalMemSize
cudaGetDeviceCount
cudaGetDeviceProperties
calloc host_output
malloc input
malloc coeff
generateRandomData

FDTD on 128 x 128 x 128 volume with symmetric filter radius 4 for 5 timesteps...

fdtdReference...
calloc intermediate
Host FDTD loop
t = 0
t = 1
t = 2
t = 3
t = 4

fdtdReference complete
calloc device_output
fdtdGPU...
cudaGetDeviceCount
cudaSetDevice (device 0)
cudaMalloc bufferOut
cudaMalloc bufferIn
set block size to 16x16
cudaMemcpy (HostToDevice) bufferIn
cudaMemcpy (HostToDevice) bufferOut
cudaMemcpyToSymbol (HostToDevice) stencil
cudaEventCreate
cudaEventCreate
GPU FDTD loop
t = 0 launch kernel
t = 1 launch kernel
t = 2 launch kernel
t = 3 launch kernel
t = 4 launch kernel

cudaThreadSynchronize
cudaMemcpy (DeviceToHost)
cudaEventElapsedTime

FDTD3d, Throughput = 90619.4714 MPoints/s, Time = 0.00002 s, Size = 1966080 Points, NumDevsUsed = 1, Blocksize = 256

cudaThreadExit
fdtdGPU complete

CompareData (tolerance 0.000100)...
Data error at point (4,4,4) 0.073752 instead of 10.781720

FAILED


!!! Error # 0 at line 48 , in file src/FDTD3d.cpp !!!

Exiting...
-----------------------------------------------------------

[cjb]

Originally Posted by gio___

FDTD3d, Throughput = 90619.4714 MPoints/s, Time = 0.00002 s, Size = 1966080 Points, NumDevsUsed = 1, Blocksize = 256

cudaThreadExit
fdtdGPU complete

CompareData (tolerance 0.000100)...
Data error at point (4,4,4) 0.073752 instead of 10.781720

FAILED


!!! Error # 0 at line 48 , in file src/FDTD3d.cpp !!!

Exiting...
-----------------------------------------------------------

Well, crap. OK, I like it less (cause it's less like what I do), but can you try the matrixMul example?

[gio___]

Originally Posted by cjb

Well, crap. OK, I like it less (cause it's less like what I do), but can you try the matrixMul example?

[ matrixMul ]
./matrixMul Starting...

Device 0: "GeForce GTX 260M" with Compute 1.1 capability

Using Matrix Sizes: A(80 x 160), B(80 x 80), C(80 x 160)

Run Kernels...

matrixMul, Throughput = 47.7019 GFlop/s, Time = 0.00004 s, Size = 2048000 Ops, NumDevsUsed = 1, Workgroup = 256

Check against Host computation...

PASSED

[cjb]

Originally Posted by gio___

matrixMul, Throughput = 47.7019 GFlop/s, Time = 0.00004 s, Size = 2048000 Ops, NumDevsUsed = 1, Workgroup = 256

Check against Host computation...

PASSED

Cool. As far as I can tell, that's a good number (better than I get from a desktop gt 240, and about half what I get from a c1070). My Alienware is reporting about 19.8. Maybe I'll have to try gentoo.... thanks for the effort!

[gio___]

Originally Posted by cjb

Cool. As far as I can tell, that's a good number (better than I get from a desktop gt 240, and about half what I get from a c1070). My Alienware is reporting about 19.8. Maybe I'll have to try gentoo.... thanks for the effort!

You're welcome
I use only the CUDA volume rendering stuff and I think there was a marked improvement in performance from 256.29 drivers with the PowerMizer/ACPI fix.
In any case in my experience, compiz and compositing in general significantly reduces the the video card performance... so I can live without some transparency if this leads to better video performace

If you need some configuration files such as mine kernel .config for gentoo just ask...

Gio

[gio___]

When I boot the laptop with AC power and I continue using it on battery mode Xorg crashes and that's is the error reported in dmesg:

NVRM: Xid (0002:00): 6, PE0001

NOTE: ACPI change correctly it's status from AC to Battery but nvidia-setting don't switch.

It's WM indipendent, I have the same crash with Gnome or Awesome WM.