Has anyone seen this odd Device to Host Transfer issue (GTX 295) Device to Host speed crippled on Co

Forum regular here, who was referred to this section of the forum by nVidia customer care in an attempt to narrow down or resolve a strange issue.

Have any of you CUDA aficionados had ever seen an issue whereby on cold boots a GTX 295 Single PCB can potentially have poor CUDA performance and suffer from a general degradation in gaming performance. On these cold boots CUDA-Z measures a drastic reduction in Device to Host and Host to Device Transfer Speeds.

Here are screenshots of CUDA-Z which I hope explain my findings in more detail.

Here we see on an occasional COLD Boot the GTX 295 has horrid performance

Here on a usual warm boot the GTX 295 is performing adequately

Has anybody seen this issue before? If so what was the cause and how did you resolve it?

The only way I have found to resolve this issue is to power the PC off (not reboot) and leave it for 20seconds before powering it back on. Rebooting sets the GTX 295 into a “half crippled mode” where it is operating @ 27xxMB/s as opposed to the 48xx MB/s

System Specifications:

Core2 Quad Extreme QX9650
8GB of DDR3 @ 1333Mhz CL9
ASUS P5E3 (X38) Motherboard with latest BIOS 1503
nVidia Geforce GTX 295 Single PCB edition.
Creative X-Fi Fatail1ty Sound Card.
2x 80GB X-25M SSD’s in RAID0
1WD Velociraptor 300GB
Enermax Infiniti 720W PSU
Windows 7 Ultimate x64

Any advice you have to offer would be most appreciated.

Thanks

John

Anyone?

Anyone?

Surely someone must have any ideas?!?
nVidia TSD are not being very helpful, they suggest it COULD be a vBIOS issue, however they do not have any updated vBIOS for me as they say only the manufacturer will have the vBIOS. Sadly it is a reference board made by a company which has gone into liquidation (BFG) :(
It is a Single PCB GTX 295 with VGA BIOS revisions 62.00.6C.00.01 and 62.00.6C.00.02.
Thanks
John

Surely someone must have any ideas?!?
nVidia TSD are not being very helpful, they suggest it COULD be a vBIOS issue, however they do not have any updated vBIOS for me as they say only the manufacturer will have the vBIOS. Sadly it is a reference board made by a company which has gone into liquidation (BFG) :(
It is a Single PCB GTX 295 with VGA BIOS revisions 62.00.6C.00.01 and 62.00.6C.00.02.
Thanks
John

John, it could be a memory partitioning issue ( bad implementation of systolic array method ), although you might wanna check the following report and/or see what CUDA is reporting with the suggested driver.

For the chronicle, I wouldn’t touch any driver released within the last year.

V801 GeForce GTX 295 896MB/DDR3 62.00.6C.00.01/02.

295 user reference

Memory Information

------------------

	Total Global: 896 MB

	Shared Per Block: 16 KB

	Pitch: 256 KB

	Total Constant: 64 KB

	Texture Alignment: 256

	GPU Overlap: Yes

John, it could be a memory partitioning issue ( bad implementation of systolic array method ), although you might wanna check the following report and/or see what CUDA is reporting with the suggested driver.

For the chronicle, I wouldn’t touch any driver released within the last year.

V801 GeForce GTX 295 896MB/DDR3 62.00.6C.00.01/02.

295 user reference

Memory Information

------------------

	Total Global: 896 MB

	Shared Per Block: 16 KB

	Pitch: 256 KB

	Total Constant: 64 KB

	Texture Alignment: 256

	GPU Overlap: Yes

Thank you for your suggestion xXiAniXx

I have the same VGA Bios in that report :)

I understand what you are saying about not touching any of the new drivers, but sadly a lot of my games work much better in SLI using 260.63 than they ever have done :(

(Metro 2033 and GTA IV)

I have ran the report and you are right, the memory does seem to be read or partitioned differently.

[codebox]Memory Information


Total Global: 869.375 MB

Shared Per Block: 16 KB

Pitch: 2.09715e+06 KB

Total Constant: 64 KB

Texture Alignment: 256

GPU Overlap: Yes[/codebox]

Any ideas?

Thanks

John

Thank you for your suggestion xXiAniXx

I have the same VGA Bios in that report :)

I understand what you are saying about not touching any of the new drivers, but sadly a lot of my games work much better in SLI using 260.63 than they ever have done :(

(Metro 2033 and GTA IV)

I have ran the report and you are right, the memory does seem to be read or partitioned differently.

[codebox]Memory Information


Total Global: 869.375 MB

Shared Per Block: 16 KB

Pitch: 2.09715e+06 KB

Total Constant: 64 KB

Texture Alignment: 256

GPU Overlap: Yes[/codebox]

Any ideas?

Thanks

John

Well, I do have few idea, alas you know problem like the one in Metro2033/BFBC2 are pretty much bound in how the application call for the creation of the DX interface.
Unless the developer of such game split the application, or nVidia find a way to handle the D3D_FEATURE_LEVEL without introducing tenthousend bug, the issue will persist.

Well, I do have few idea, alas you know problem like the one in Metro2033/BFBC2 are pretty much bound in how the application call for the creation of the DX interface.
Unless the developer of such game split the application, or nVidia find a way to handle the D3D_FEATURE_LEVEL without introducing tenthousend bug, the issue will persist.

So are you saying that in order for me to have good performance in Metro 2033 I have to put up with this odd bug?

And a loss of some of my VRAM?

----------

Total Global: 869.375 MB

Shared Per Block: 16 KB

Pitch: 2.09715e+06 KB

Total Constant: 64 KB

Texture Alignment: 256

GPU Overlap: Yes

Notice how 869MB is reported as opposed to 896MB :(

Also the Pitch is different to yous too!

I have reported this in the 260.63 driver feedback thread…

John

So are you saying that in order for me to have good performance in Metro 2033 I have to put up with this odd bug?

And a loss of some of my VRAM?

----------

Total Global: 869.375 MB

Shared Per Block: 16 KB

Pitch: 2.09715e+06 KB

Total Constant: 64 KB

Texture Alignment: 256

GPU Overlap: Yes

Notice how 869MB is reported as opposed to 896MB :(

Also the Pitch is different to yous too!

I have reported this in the 260.63 driver feedback thread…

John

I live on the other hemisphere from where the nVidia office stand, otherwise, you know… I would been there to kick some ass - night&day.

I live on the other hemisphere from where the nVidia office stand, otherwise, you know… I would been there to kick some ass - night&day.

SOLVED

I replaced the motherboard with an ASUS Rampage Extreme (X48)! The issue appears to have gone away. All other components remain the same, just the motherboard swapped.

John

SOLVED

I replaced the motherboard with an ASUS Rampage Extreme (X48)! The issue appears to have gone away. All other components remain the same, just the motherboard swapped.

John