Hi!
Does anyone know how many global memory partitions there are on a GTX 260? 8 ?
Anyways I’m gonna experiment a bit so maybe i will be able to figure out, but it would be nice with a second source.
Thanks
J
GTX 260 has 448-bit interface, 448 = 64 x 7, so it should has 7 partitions (channels)
,the same as GTX295
This is a Quadro FX 4800, with a 384 bits bus (?)… That would yield 384/64 = 6 partitions…
But i read
“Just as shared memory is divided into 16 banks of 32-bit width, global memory is
divided into either 6 partitions (on 8- and 9-series GPUs) or 8 partitions (on 200-
and 10-series GPUs) of 256-byte width.”
- Optimizing Matrix Transpose in CUDA, whitepaper
So it’s 8 ? :) I’m a bit confused right now…
Divide the bus width by 64 to get the number of partitions.
512 -->8
448 -->7
384 -->6
Thanks!
I had to test a bit :)
6 it is. notice the dips in transfer at multiples of 6 ( 6, 12, 18, 24 etc).
Datasize: 8192 KB
Data transfer: 18.869
--------------------------------------------------------------
Partitions 2
Datasize: 8192 KB
Data transfer: 17.241
--------------------------------------------------------------
Partitions 3
Datasize: 8191 KB
Data transfer: 12.835
--------------------------------------------------------------
Partitions 4
Datasize: 8192 KB
Data transfer: 18.506
--------------------------------------------------------------
Partitions 5
Datasize: 8191 KB
Data transfer: 18.853
--------------------------------------------------------------
Partitions 6
Datasize: 8191 KB
Data transfer: 8.032
--------------------------------------------------------------
Partitions 7
Datasize: 8191 KB
Data transfer: 18.777
--------------------------------------------------------------
Partitions 8
Datasize: 8192 KB
Data transfer: 17.731
--------------------------------------------------------------
Partitions 9
Datasize: 8190 KB
Data transfer: 12.838
--------------------------------------------------------------
Partitions 10
Datasize: 8190 KB
Data transfer: 16.361
--------------------------------------------------------------
Partitions 11
Datasize: 8189 KB
Data transfer: 18.830
--------------------------------------------------------------
Partitions 12
Datasize: 8190 KB
Data transfer: 8.276
--------------------------------------------------------------
Partitions 13
Datasize: 8190 KB
Data transfer: 19.088
--------------------------------------------------------------
Partitions 14
Datasize: 8190 KB
Data transfer: 15.073
--------------------------------------------------------------
Partitions 15
Datasize: 8190 KB
Data transfer: 12.251
--------------------------------------------------------------
Partitions 16
Datasize: 8192 KB
Data transfer: 16.460
--------------------------------------------------------------
Partitions 17
Datasize: 8189 KB
Data transfer: 16.784
--------------------------------------------------------------
Partitions 18
Datasize: 8190 KB
Data transfer: 8.164
--------------------------------------------------------------
Partitions 19
Datasize: 8189 KB
Data transfer: 17.585
--------------------------------------------------------------
Partitions 20
Datasize: 8190 KB
Data transfer: 13.689
--------------------------------------------------------------
Partitions 21
Datasize: 8190 KB
Data transfer: 12.010
--------------------------------------------------------------
Partitions 22
Datasize: 8189 KB
Data transfer: 12.918
--------------------------------------------------------------
Partitions 23
Datasize: 8188 KB
Data transfer: 19.160
--------------------------------------------------------------
Partitions 24
Datasize: 8190 KB
Data transfer: 8.322
--------------------------------------------------------------
Can anybody think of a way to determine the number of partitions programmatically for an unknown card? I notice that the device properties structure does not contain the width of the memory bus although it does contain the total amount of memory.
You could of course write some code that tests for which number of partitions that the data transfer rate drops. But that’s a bit brute force and unsophisticated…
I guess this can help your code determine how to best avoid partition camping on different hardware.