Help for buying CUDA low-end workstation

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everybody,

I started my PhD this month on a project about numerical optimization and machine learning with CUDA/OpenCL. I have around $1500 to spend for my desktop computer and I was asking myself what should I buy. Can you please tell me which components would you buy if you were in my position? GPU? CPU? RAM?

Thanks a lot for your help.
den3b

2

If the goal is to get as much computer as possible for $1500, I would suggest a Core i7-930 with 6 GB of memory (triple channel), GTX 470, and a power supply capable of driving 2 cards, and maybe a modestly sized SSD. Then, as needed, you can decide what to upgrade during your Ph.D: more RAM, extra GPU, more storage.

3

If the goal is to get as much computer as possible for $1500, I would suggest a Core i7-930 with 6 GB of memory (triple channel), GTX 470, and a power supply capable of driving 2 cards, and maybe a modestly sized SSD. Then, as needed, you can decide what to upgrade during your Ph.D: more RAM, extra GPU, more storage.

4

$1500 can buy you a lot of computer, especially if you build it yourself.

i7 930 $285
Decent motherboard with support for 24GB of ram and SLI (think of the future) $180
750W power supply $125
6GB ram $160
GTX 470 $315 if you look hard, $350 if you don’t.
CD and HD $200

So for $1300, you can get a pretty sweet computer. This is pretty much what I have, except for HDs and CD/blu-ray.

5

$1500 can buy you a lot of computer, especially if you build it yourself.

i7 930 $285
Decent motherboard with support for 24GB of ram and SLI (think of the future) $180
750W power supply $125
6GB ram $160
GTX 470 $315 if you look hard, $350 if you don’t.
CD and HD $200

So for $1300, you can get a pretty sweet computer. This is pretty much what I have, except for HDs and CD/blu-ray.

6

I was recently in the same situation-- This is more or less what I went with:

2X Western Digital Caviar Black WD1002FAEX 1TB 7200 RPM SATA 6.0Gb/s 3.5" Internal HHD
$180

ASUS M4A89TD PRO/USB3 AM3 AMD 890FX SATA 6Gb/s USB 3.0 ATX AMD Motherboard
$175

16GB (4 x 4GB) 240-Pin DDR3 SDRAM DDR3 1333
$500

AMD Phenom II X6 1090T Black Edition
$300

Computer Case + 750W Power Supply + DVD burner ~=
$200

GeForce GTX 470 =
$300

This is slightly over your limit but if you’re a computational guy like me:
2 HDD’s for large project files, backup purposes and redundancy.
The MB has fast bus speed and supports fast and high capacity memory.
I assume your problem set is large (otherwise, why bother with HPC solutions?), in which case I easily use up 16GB of memory. This is good for running two OS’s and many programs at once. With only 6GB as suggested, I wouldn’t have much memory left to run my code…
I like the 6 core AMD 1090T so I can run linux on 4 cores and also run windows with VM player with 2 cores. Also an excellent choice for use with other HPC solutions (MPI, OpenMP,…). I tend to like AMD more than Intel since I think you get more bang for your buck.
I have a large tower and power supply for nice air circulation and future expansion.
As far as a CUDA card goes, getting a new 400 series with the Fermi arch is the way to go.

Just my thoughts…

7

I was recently in the same situation-- This is more or less what I went with:

2X Western Digital Caviar Black WD1002FAEX 1TB 7200 RPM SATA 6.0Gb/s 3.5" Internal HHD
$180

ASUS M4A89TD PRO/USB3 AM3 AMD 890FX SATA 6Gb/s USB 3.0 ATX AMD Motherboard
$175

16GB (4 x 4GB) 240-Pin DDR3 SDRAM DDR3 1333
$500

AMD Phenom II X6 1090T Black Edition
$300

Computer Case + 750W Power Supply + DVD burner ~=
$200

GeForce GTX 470 =
$300

This is slightly over your limit but if you’re a computational guy like me:
2 HDD’s for large project files, backup purposes and redundancy.
The MB has fast bus speed and supports fast and high capacity memory.
I assume your problem set is large (otherwise, why bother with HPC solutions?), in which case I easily use up 16GB of memory. This is good for running two OS’s and many programs at once. With only 6GB as suggested, I wouldn’t have much memory left to run my code…
I like the 6 core AMD 1090T so I can run linux on 4 cores and also run windows with VM player with 2 cores. Also an excellent choice for use with other HPC solutions (MPI, OpenMP,…). I tend to like AMD more than Intel since I think you get more bang for your buck.
I have a large tower and power supply for nice air circulation and future expansion.
As far as a CUDA card goes, getting a new 400 series with the Fermi arch is the way to go.

Just my thoughts…

8

If you swap to the MSI 980i-G65, you save $20, and get onboard display (very very very useful for CUDA!) and you can scale up to 3 GPU on top of that!

The M4A89TD will only support 2 GPU, and you need to use one of them for display… so it’s really limited to 1 GPU.

9

If you swap to the MSI 980i-G65, you save $20, and get onboard display (very very very useful for CUDA!) and you can scale up to 3 GPU on top of that!

The M4A89TD will only support 2 GPU, and you need to use one of them for display… so it’s really limited to 1 GPU.

10

I must admit I dedicate 1 slot for display and this is far from ideal in a GPGPU environment. My machine is more geared toward using CPU parallelization models. These multiple slots would be an excellent choice for den3b.

11

I must admit I dedicate 1 slot for display and this is far from ideal in a GPGPU environment. My machine is more geared toward using CPU parallelization models. These multiple slots would be an excellent choice for den3b.

12

Thanks a lot guys,

this forum rocks!

I will build my configuration based on your suggestions :). I’ll let you know before final decision :D

13

Thanks a lot guys,

this forum rocks!

I will build my configuration based on your suggestions :). I’ll let you know before final decision :D

14

Hey guys,

what about this configuration. It’s just around $1500 here in Hong Kong.

CPU
Intel Core i7-870 2.93GHz 8 MB LGA1156

MotherBoard
EVGA P55 Classified 200 160-LF-E659-KR (3way SLI support)

RAM
2x D3-1333 4GB Samsung CL RAM (8 GB Total)

Graphic Card
MSI nVidia GeForce GTX 470 1280 MB

Hard Disks
1TB SATA2 Seagate HD
40 GB Intel X25-V SSD

Other
DVD Reader
Acbel R9 900W 80 Plus Gold Power Supply
Gaming Case ATX

do you think it is worth it, or I would be better off with an AMD solution?

Thanks a lot,
den3b

15

Hey guys,

what about this configuration. It’s just around $1500 here in Hong Kong.

CPU
Intel Core i7-870 2.93GHz 8 MB LGA1156

MotherBoard
EVGA P55 Classified 200 160-LF-E659-KR (3way SLI support)

RAM
2x D3-1333 4GB Samsung CL RAM (8 GB Total)

Graphic Card
MSI nVidia GeForce GTX 470 1280 MB

Hard Disks
1TB SATA2 Seagate HD
40 GB Intel X25-V SSD

Other
DVD Reader
Acbel R9 900W 80 Plus Gold Power Supply
Gaming Case ATX

do you think it is worth it, or I would be better off with an AMD solution?

Thanks a lot,
den3b

16

I should note that the socket LGA1156 has limited bandwidth between CPU and PCI-Express bus. Unlike the LGA1366, a PCI-Express x16 2.0 link is built right into LGA1156 CPUs along with a slower second bus to the rest of the peripherals. I don’t know how important multi-GPU is for your future expansion, but it is possible that adding a second GPU will cut your bandwidth in half with this configuration. An LGA1366 CPU has enough bandwidth to run two GPUs full speed with some additional head-room. (The triple channel RAM interface on LGA1366 also helps.)

17

I should note that the socket LGA1156 has limited bandwidth between CPU and PCI-Express bus. Unlike the LGA1366, a PCI-Express x16 2.0 link is built right into LGA1156 CPUs along with a slower second bus to the rest of the peripherals. I don’t know how important multi-GPU is for your future expansion, but it is possible that adding a second GPU will cut your bandwidth in half with this configuration. An LGA1366 CPU has enough bandwidth to run two GPUs full speed with some additional head-room. (The triple channel RAM interface on LGA1366 also helps.)

18

Uh… that’s quite a point, thanks seibert. Does anybody know if a Phenom II X6 has the same bus/bandwidth problem between CPU and multi-GPUs? If not, any idea how much will the lack of triple channel RAM interface on Phenom II reduce performances when using multiple GPUs?

Thanks

den3b

19

Uh… that’s quite a point, thanks seibert. Does anybody know if a Phenom II X6 has the same bus/bandwidth problem between CPU and multi-GPUs? If not, any idea how much will the lack of triple channel RAM interface on Phenom II reduce performances when using multiple GPUs?

Thanks

den3b

20

Socket AM3 uses a HyperTransport interface to the chipset, which has plenty of bandwidth. I’ve had good success with motherboards using the 790FX chipset, though that has been superseded by the 890FX now. The Phenom II X6 supports 4 GHz HT, which gives you 16 GB/sec to the chipset (assuming your motherboard/chipset also supports 4 GHz). You do see a slight drop in PCI-E bandwidth switching from AMD to Intel LGA1366 with triple channel. The memory bandwidth provided by the triple channel interface is high enough that most Core i7 systems have nearly the same performance doing host-to-device transfers from pinned and pageable host memory. (Transfers from pageable memory require an intermediate copy to a pinned buffer in the CUDA driver and generally run around half-speed.) AMD isn’t bad though. If I recall correctly, in practice I saw something like 5 GB/sec for cards in my AMD system and 6 GB/sec for the Core i7.

I think the AMD option is fine if the Core i7 LGA1366 is too much.