It is safe to use GeForce 1070 and Titan Xp with a 750W PSU?
How much Watts a PSU must have for running both GPUs?
What other components (CPU, DRAM, mass storage) are in the systemm?
There is no hard rule for PSU sizing. It depends on your usage patterns (do you expect to put 100% load on both GPUs simultaneously?) and on what likelihood of flakiness (up to sudden machine resets) you are willing to tolerate.
My empirically determined rule of thumb for rock-solid operation is that the nominal power of all system components combined should be < 60% of the nominal power of the PSU. Others might claim that this is too conservative.
Using a low-ball estimate of 120W for all non-GPU components in your system would already but us at 520W of combined nominal power, which would require >= 850W PSU under my rule of thumb.
Hi, thanks for the answer. The components are:
- CPU: Intel(R) Core™ i7-6700K CPU @ 4.00GHz;
- DRAM: 4x Hyperx 8GB 2400MHz (32GB in total);
- Mass Storage: Seagate 1TB SATA 3.5.
The both GPU will be used to machine learning, and probably will use 100%. But the others components will be slightly used.
You think that is safe to perform a performance test with the 2 GPU, to test PSU?
Intel lists the CPU with a TDP of 91W. 32 GB of DDR4 @ 0.4W/GB = 13W. HDD ~ 6W. Miscellaneous motherboard components combined ~ 10W. So that’s 120W altogether, which is what I had assumed in my earlier post. The Titan XP is specified with 250W, and the GTX 1070 with 150W. Grand total of 520W.
Machine learning typically means ~ 100% GPU utilization for many hours at a time, i.e. continuous heavy load.
By my rule of thumb that suggests an 850W PSU. 80 PLUS Platinum rated would be best: high efficiency, less wasted energy, usually built from higher-quality components. If you cannot afford an 80 PLUS Platinum rated power supply, I would recommend an 80 PLUS Gold rated one as the minimum for a workstation-class machine.
My personal experience: If the PSU is extremely overloaded, you will likely experience a hard machine reset most times you start a CUDA program that uses both GPUs. If you simply run at very high load, the PSU will run hot, and will die in as little as six months. All electronic components age, and the speed of many of the underlying processes roughly doubles for every 10 deg Celsius increase in temperature, by Arrhenius’ Rule.
If you ever work with a large pool of machines, you will find that the component mortality is highest for PSUs, followed by DRAM. At least that is what I have seen.