kolonel
February 10, 2010, 11:16pm
1
Hi,
I have a GTX 280 card and one unit of Tesla S1070 connected to my computer (so, there are 5 devices), I wanted to know how much RAM I should install on the host computer to have a good efficiency for having 4 GPUs of the S1070 work in parallel?
Right now I have 4GB of RAM, DDR2.
Thanks,
LSChien
February 11, 2010, 12:17am
2
Hi,
I have a GTX 280 card and one unit of Tesla S1070 connected to my computer (so, there are 5 devices), I wanted to know how much RAM I should install on the host computer to have a good efficiency for having 4 GPUs of the S1070 work in parallel?
Right now I have 4GB of RAM, DDR2.
TeslaS1070 has 4 GPUs, each can manipulate 4GB GDDR3. If you want to build a multi-GPU application and exhaust all memory in S1070, then you should have the same amount of host RAM. Hence 16GB host RAM is good.
do you connect to all 4 GPUs of S1070, or just two among them?
(S1070 has two PCI adapter, each adapter control two GPUs)
kolonel
February 11, 2010, 12:28am
3
Thanks for the reply,
I have connected all 4 GPUs…
if having 16GB of RAM makes a significant difference in the speed of computations with s1070?!
is there any reference (link or document) for the minimum memory requirement for s1070?
TeslaS1070 has 4 GPUs, each can manipulate 4GB GDDR3. If you want to build a multi-GPU application and exhaust all memory in S1070, then you should have the same amount of host RAM. Hence 16GB host RAM is good.
do you connect to all 4 GPUs of S1070, or just two among them?
(S1070 has two PCI adapter, each adapter control two GPUs)
LSChien
February 11, 2010, 8:58am
4
It is not necessary that host memory is larger than GPU memory.
It depends on your application.
For example, consider a model problem, matrix multiplcation C = A * B (double precision),
A, B and C are square matrices with size 32K x 32K (each matrix has 8GByte).
To store three matrices, you need 24GB in your host memory, and you cannot use one GPU of S1070
to compute C = A*B. You need to divide matrix C to several tiles, and each GPU can compute one tile of C.
In this case, you need large host memory.
kolonel
February 11, 2010, 6:02pm
5
I am doing a large scale computation, so I need large amount of memory,
but my host computer’s os is Windows XP64, and it is not supporting more than 4GB of RAM, right? even if I install 16GB of RAM it is not usable…
It is not necessary that host memory is larger than GPU memory.
It depends on your application.
For example, consider a model problem, matrix multiplcation C = A * B (double precision),
A, B and C are square matrices with size 32K x 32K (each matrix has 8GByte).
To store three matrices, you need 24GB in your host memory, and you cannot use one GPU of S1070
to compute C = A*B. You need to divide matrix C to several tiles, and each GPU can compute one tile of C.
In this case, you need large host memory.
CapJo
February 11, 2010, 6:07pm
6
I am doing a large scale computation, so I need large amount of memory,
but my host computer’s os is Windows XP64, and it is not supporting more than 4GB of RAM, right? even if I install 16GB of RAM it is not usable…
From TechNet
Benefits of 64-bit Windows
Additional Performance and Scalability
Large memory support. Supports up to 16 GB of RAM and 16 TB of virtual memory, enabling applications to run faster when working with large data sets. Applications can preload substantially more data into virtual memory, allowing rapid access by the Intel Itanium processor. This reduces the time for loading data into virtual memory or seeking, reading and writing to data storage devices, thus making applications run faster and more efficiently.
I would suggest to switch to Windows 7 64 bit edition. Debugging on hardware is only possible with Vista and Windows 7. Emulation Support will be discontinued with CUDA 3.0.
LSChien
February 12, 2010, 12:45am
7
I am doing a large scale computation, so I need large amount of memory,
but my host computer’s os is Windows XP64, and it is not supporting more than 4GB of RAM, right? even if I install 16GB of RAM it is not usable…
I use winxp64 also, my host memory is 16GB.
However if you want to use Nexus (debugger and profiler integrated into vs2008 on vista and window7),
please install vista64 or window7-64.
Xvv
February 12, 2010, 3:17pm
8
I am doing a large scale computation, so I need large amount of memory,
but my host computer’s os is Windows XP64, and it is not supporting more than 4GB of RAM, right? even if I install 16GB of RAM it is not usable…
I regularly setup XP64 computers with 24GB of RAM. XP64 supports 128GB RAM (cf. msdn Memory Limits for Windows Releases ).
Be careful which version of win7x64 you select (if you go that way): apparently there are pretty severe limitations for “Starter” (2GB max), “Home Basic” (4GB max) and “Home Premium” (16GB max)).
Note: If your system has less memory than the S1070, you can always load/unload data to each gpu “sequentially”, so it does work. How badly your performances will be impacted depends on your specific algorithm.