Multi GPU , how to increase the training speed?
Assume that you have enough budget, what's the best way to increase the training speed from hardware perspective? SLI doesn't seem to work as addressed in this thread, will NVlink work? Is it that only the VRAM from those multi GPUs are combined for the training?
To give a though to your questions, here are some anwsers although not necessarily THE answers.
Assume that you have enough budget, what's the best way to increase the training speed from hardware perspective?
- What does enough budget mean, is enough: a dollar, a thousand dollars, unlimited dollars? It is probably best to go for one very powerful NVIDIA GPU with tons of VRAM instead of multiple weaker ish ones. This is due to possible PCIe lane, compatibly, power consumption and cooling problems.
SLI doesn't seem to work as addressed in this thread, will NVlink work? Is it that only the VRAM from those multi GPUs are combined for the training?
- When using NVlink you should be able to utilize the two connected GPU's simultaneously. It should not only combine VRAM but also increase the processing speed by the expected 100 percent when using two identical cards. (Although this can be less or more in the practical reality depending on what you are processing at the moment.)
Hopefully you can plan a bit further now, if you need any more information, feel free to ask.
Please close this issue.