CUDA bandwidth test Shmoo Anyone else got some results

Hi all,

Does anyone have any results of a shmoo test (pinned or paged) that they can attach to a post please? I’m having some really slow issues writing back using cudaMemCpy, but I suspect this has more to do with not using pinned memory. Just want to check my ULi PCI Express chipset + 8800GT are in the same ballpark as others.

Thanks a lot in advance

Shmoo Mode

Host to Device Bandwidth for Pinned memory

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  5347.1

Shmoo Mode

Device to Host Bandwidth for Pinned memory

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  4590.3

Shmoo Mode

Device to Device Bandwidth

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  51735.3

&&&& Test PASSED

Here’s the important stuff from some results. I don’t want to spam the board with the whole thing unless it’s necessary. PM me if you’d like the rest, I guess. It’s an 8800 GTS in a Tyan server board. We have some standard desktop motherboards too, I’ll try to run tests on those as well. They’re slower though, I think.

Here are the pagaeable results, cropped:

Shmoo Mode

Host to Device Bandwidth for Pageable memory

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  1607.6

Shmoo Mode

Device to Host Bandwidth for Pageable memory

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  1464.6

Shmoo Mode

Device to Device Bandwidth

.................................................................................

Transfer Size (Bytes)	Bandwidth(MB/s)

 67186688  51714.5

&&&& Test PASSED

That looks much more like PCI Express v2 hence the speedup, infact wasnt it 4GB/s per direction.

Dont happen to have the results of a full Shmoo pinned in both directions? I’ll PM for in, in that case.

Here are the results on my machine (780i MB, 8800 GTS 512MB GPU, DDR 8000 RAM). Host <–> device rates are reported to be slightly faster on Intel P35 chipsets.
bandwidth_test_shmoo_780i.gz (1.74 KB)

Thanks a bunch for these guys.

I just compared to mine :

Shmoo Mode
Device to Host Bandwidth for Pinned memory

.
Transfer Size (Bytes) Bandwidth(MB/s)
1024 0.6
2048 89.1
3072 129.9
4096 167.9
5120 203.2
6144 237.0
7168 259.8
8192 302.7
9216 331.5

I think I’ve identified my problem here… time to look at another chipset sometime, me thinks.

For the record, its the ULi M1689 - the one with AGP and PCI-E v1 support. Shame, as this doesnt show up in gaming, though when did a game’s graphics rendering feed back much to the system?

Here are a couple of complete tests. Didn’t realize you could attach files. Duh!

These are 8800 GTS cards. One is in a Tyan S5397 board, and the other in a Gigabyte P35-DS4.
shmoo_pinned_Tyan_S5397_8800GTS.txt.zip (1.88 KB)
shmoo_pinned_Gigabyte_P35_DS4_8800GTS.txt.zip (1.86 KB)

Sorry to dig up old posts, but I was wondering if anybody knew if the most significant limiting factor for the smaller transfers as illustrated in this post with the shmoo test is the pci bus, host cpu or gpu memory interface.

Matt

Probaby the CPU to handle the overhead of set-up. Hmm, i forgot to run shmoo when I did the downclocking of my CPU and FSB. (Btw, it’s also possible to change PCI-E frequency on my board, to see what effect that has.)

[bandwidthTest] starting…

./bandwidthTest Starting…

Running on…

Device 0: GeForce GTX 590

Quick Mode

Host to Device Bandwidth, 1 Device(s), Paged memory

Transfer Size (Bytes) Bandwidth(MB/s)

33554432 5440.0

Device to Host Bandwidth, 1 Device(s), Paged memory

Transfer Size (Bytes) Bandwidth(MB/s)

33554432 4957.8

Device to Device Bandwidth, 1 Device(s)

Transfer Size (Bytes) Bandwidth(MB/s)

33554432 113736.5

[bandwidthTest] test results…

PASSED