CUDA on external NVIDIA GPU [closed]

Question

Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.

---

This question does not appear to be about a specific programming problem, a software algorithm, or software tools primarily used by programmers. If you believe the question would be on-topic on another Stack Exchange site, you can leave a comment to explain where the question may be able to be answered.

Closed 9 years ago.

Improve this question

My system has an AMD A6 processor with Radeon 7470M+HD6520G dual graphics. So it is a complete AMD system! I run AMD APP OpenCL SDK on this efficiently. I am interested to use CUDA on a more powerful NVIDIA GPU. My question is,

• If I buy an NVIDIA GPU with PCI express enabled, can I connect it to my system and run CUDA?

• How much bottleneck comes in the memory transfer between host and device(CPU-GPU)?

Answer 1

If I buy an NVIDIA GPU with PCI express enabled, can I connect it to my system and run CUDA?

Yes, sure! Just make sure that you have the correct PCI express slot and a correct Power Supply Unit (some high-end GPUs need an extra power line, because PCI express can provide a limited amount of current).

How much bottleneck comes in the memory transfer between host and device(CPU-GPU)?

Theoretically GPUs sitting on PCIe 2.0 x16 are capable of 16*5GT/s, that because of the error correction encoding (2 ecc bits every 10 bits, 8/10 encoding) become 16*4Gbit/s: 8GB/s.

If your systems has x16 PCIe 3.0 as the same as the GPU, the theoretical transfer rate is: 16*8GT/s. Here the error correction scheme is much more lightweight (128/130 encoding), so it's approximatively 16*8 Gbit/s = 16GB/s.

I've seen for real up to 6GB/s on PCIe 2.0 systems - never touched one with PCIe 3.0 actually.