This may not be the WU fault- 66xx's [NVidia-8400GS]

[bruce]

(I'm not up-to-speed on the support details for specific hardware so I can only ask "stupid" questions.) Does NV say that the 8400GS have full CUDA support? Is OpenCL support planned or will that support be lacking?

Future support (either in GPU2 or in GPU3) depends directly on what NV does. At the very least, Species=0 is not a good sign.

[new08]

Not sure about that Bruce. it's an oldish card now- so I doubt any new moves.
Drivers are , however, listed as CUDA 2.2 capable and as nVidia are pros- that is what it means. Your question asks for whether this is backwards compatible with hardware capabilities- I don't know, tbh.
If we knew what rigs had returned 66xx units we would be a little nearer resolution but that query is unlikely to run anywhere I can think of! - [Bar -PG ;~/ ]
In the broad view- I would see my quest as dweebish- except for the fact that a hole opening up in the arena of GPU2/3, CUDA, older cards g80 unresolved issues- so it is interesting to pursue from an technical viewpoint, apart from optomising my gear.

[bretth603]

Sorry to jump in on this late....

I'm running a Quadro NVS 295 which is still an active product and is marketed as CUDA-capable. It uses the same chipset as the 8400GS and has exactly the same problems with FAH. It is also identified as Species=0 by FAH. I don't think it is an age or lack of driver support issue.

[PantherX]

Please see this: viewtopic.php?f=38&t=15940#p157837

[new08]

Sorry to jump in on this late....

I'm running a Quadro NVS 295 which is still an active product and is marketed as CUDA-capable. It uses the same chipset as the 8400GS and has exactly the same problems with FAH. It is also identified as Species=0 by FAH. I don't think it is an age or lack of driver support issue.

Brett- Do you weed out the 66xx units in your firewall ?-
[Overclocked GPU?]
It can [sometimes] take a an hour or more to get a working unit, but at the GPU work rate, worth waiting for!
Must be thousands of these cards lying around that could be used for folding- faster than many middling CPUs.

[bretth603]

I block work server 171.64.65.61 using a static IP route on the individual PC, but using a firewall would be easier for most ppl. So far it has prevented me from getting any GPU WUs that crash. Sometimes it can take several hours to get a WU, but it's well worth not crashing the GPU.

I'm not overclocking my GPU, especially since it is fanless. I tried a little AC-powered lab bench fan and was able to easily OC 25%, but it was too noisy for daily use.

[new08]

This card [8400] is not highest spec and though mine has its fan it is flat out and not boostable 'as is' by using software.
I added a PCI slot fan adjacent for ~$3 cost -and that helps a bit [with the case fan also ] as I o/c the card, if it remains stable for work.
I keep GPU core to temp 65c max though!
I find using CPU-Z utility helpful in showing graphed temps and memory management overhead nicely while trying different configs of card options

[new08]

ONE Thing, that hast slipped by in all this discussion, is that when the GPU client calls for work, my memory is listed as 1023 Mbytes rather than the cards actual 512.

Code: Select all

[05:34:50] - Machine ID: 2
[05:34:50] 
[05:34:50] Gpu type=2 species=0.
[05:34:50] Loaded queue successfully.
[05:34:50] Initialization complete
[05:34:50] - Preparing to get new work unit...
[05:34:50] Cleaning up work directory
[05:34:50] + Attempting to get work packet
[05:34:50] - Autosending finished units... [October 8 05:34:50 UTC]
[05:34:50] Trying to send all finished work units
[05:34:50] + No unsent completed units remaining.
[05:34:50] - Autosend completed
[05:34:50] - Will indicate memory of 1023 MB
[05:34:50] Gpu type=2 species=0.
[05:34:50] - Detect CPU. Vendor: AuthenticAMD, Family: 15, Model: 12, Stepping: 2
[05:34:50] - Connecting to assignment server
[05:34:50] Connecting to http://assign-GPU.stanford.edu:8080/
[05:34:52] Posted data.
[05:34:52] Initial: 40AB; - Successful: assigned to (171.64.65.71).

Is this the possible cause of certain work units not running OK?
The fact that memory is reported must mean it's a relevant factor, but not necessarily critical in all cases.

[toTOW]

The memory reported here is the system memory, not the GPU one. And I think this setting is not being used by the GPU servers to assign WUs.

[new08]

Yes,Ttow- I did wonder that -as I have 1Gb system memory.
This memory size is a common on GPUs now [and partly why they're so useful].
Sometimes, a little more info coming down the line to donors wouldn't go amiss however, by being in the right context!

[toTOW]

I don't think it has changed yet : the WUs are designed to run on card with 512 MB.

[bruce]

I've never seen a specific requirement for VRAM size for either ATI or nvidia_g80 or nvidia_Fermi from the programmers who would actually understand this issue. It is commonly accepted on the forum that FAH doesn't use much VRAM but rather moves data back and forth to main RAM as needed. The size of VRAM is important if you expect good gaming performance and high resolution images on a big monitor.

The SMP and Uniprocessor clients can put varying demands on main RAM, and the difference between 512K and 1024K might be important if you run either of those clients. I don't remember a case of this number being wrong before if we're talking about main RAM.

[new08]

Update on CUDA : I posted this earlier- "Drivers are , however, listed as CUDA 2.2 capable.."
Now, having found the the 'Z' utilities the latest being Cuda Z showing this result:-

Code: Select all

CUDA-Z Report
=============
Version: 0.5.95
http://cuda-z.sourceforge.net/
OS Version: Windows x86 5.1.2600 Service Pack 3

Core Information
----------------
	Name: GeForce 8400 GS
	Compute Capability: 1.1
	Clock Rate: 1852 MHz
	Multiprocessors: 1
	Warp Size: 32
	Regs Per Block: 8192
	Threads Per Block: 512
	Watchdog Enabled: Yes
	Threads Dimentions: 512 x 512 x 64
	Grid Dimentions: 65535 x 65535 x 1

Memory Information
------------------
	Total Global: 511.688 MB
	Shared Per Block: 16 KB
	Pitch: 2.09715e+06 KB
	Total Constant: 64 KB
	Texture Alignment: 256
	GPU Overlap: No

Performance Information
-----------------------
Memory Copy
	Host Pinned to Device: 46.3252 MB/s
	Host Pageable to Device: 45.293 MB/s
	Device to Host Pinned: 45.9824 MB/s
	Device to Host Pageable: 44.8545 MB/s
	Device to Device: 2748.67 MB/s
GPU Core Performance
	Single-precision Float: 29235 Mflop/s
	Double-precision Float: Not Supported
	32-bit Integer: 5856.19 Miop/s
	24-bit Integer: 29232.1 Miop/s

Generated: Wed Oct 13 22:03:06 2010

This clearly shows only CUDA 1.1 as being functional to the system, whatever the drivers may say- so could this be a factor in some GPU work units not running ? This may affect the GPU3 Console client too,which displays different errors.
NB: For those,unlike me, who may know the fine technical points- please note the figures above are for a 'PCI' version of the 8400GS- not the usual 'PCIe' version!

[bretth603]

Compute Capability: 1.1

Hmmm...Now that's interesting. Mine (Quadro NVS 295) says the same thing. I'm also using 258.96. If true, that would explain the species=0 (obviously). Good find new08!

Sadly, according to info. on the NVIDIA website that I don't remember seeing before, CUDA 1.1 is the highest my card will support. You have the same chipset (G98) so you will have the same limitation.

These are core11 WUs, though, so do we need higher than CUDA 1.1?

[PantherX]

Minimum requirement for FahCore_15:
In order to get WUs using this new core, donors will need to make sure their CUDA level is least CUDA 2.2
http://folding.typepad.com/news/2010/04 ... later.html