This machine is used for personal machine learning / deep learning projects.
Keywords: GPU 1080 Ti, waterblock, overclock, temperature, machine learning, budget build
CPU choice: i3-7100. Since Intel has this latest 7 series, why not choose the 7 series 14nm FinFET technology? Also, because machine learning / deep learning tasks are mainly operated in GPU, for CPU, two cores with four threads are totally enough. No demand on CPU overclock as well, then the choice comes to i3-7100, low price, enough performance.
GPU choice: Nvidia GTX 1080 Ti. This is the best performance best price GPU for personal machine learning tasks till now. Titan x (pascal) or Titan xp is too expensive but that extra money does not equal to the small performance gain compared to 1080 Ti. For full-load machine learning / deep learning tasks, it seems the usage of GPU is averagely higher than gaming, so the temperature will be crazily high. I made the stress test with founder edition fans, at 100% speed of the fan (manually tuned to 100%), temperature can be 62~65 degree C, sometimes to close to 68 degree C if the system ambient is hot. The other problem is manually tuned 100% fan speed is too loud like a jet going to fly high up into the sky. Therefore a watercool is a must for ML/DL applications. Lower temperature always means better performance and long life-time duration.
GPU performance with watercool system: 10.5 TOPS/s ~ 11.5 TOPS/s (operations per second), it depends on the main GPU frequency setups. Can be overclocked to quite stable 1760MHz ~ 2050MHz (cannot get more than 2050MHz). Temperature at idle: 24~28 degree C. Temperature at full-load with ML/DL applications with OC 2050MHz: 42~45 degree C (depends on ambient, can be lower, also depends on watercool fans, I didn't use many of them.)
Motherboard choice: H270. No need to CPU overclock, GPU overclock can be done by Nvidia-settings (with cool-bits setups). This works perfectly with i3-7100. Price of this motherboard seems to guarantee a robust performance, I mean no system glitches from motherboard at least. Not a big fan of doing motherboard BIOS, so didn't choose any 170 based motherboard. For long term, 1080 Ti will be taken out from this machine, and this machine will become some other personal project's server. So only 1 pcie slot does not matter. H270 will be enough for other maybe raspberry pi or similar thing related home projects.
Memory choice: 2*8GB. For 11GB GPU memory, total 16GB system memory is quite enough.
Storage choice: low budge SSD. 240GB is enough. HDD using 3TB to storage big data.
System power choice: 650W. This is enough for current system build. Current power consumption will be no larger than 500W, including full-load GPU.
Case choice: Manta. Just aim for sexy looking, nothing else.
Other choices: two monitor is a must, good keyboard for coding is a must, good headphones to enjoy music and videos is a must.
Long term upgrade plan: Expand the system, or say, taking the 1080 Ti out, go to four 1080 Ti all watercooled, with maybe x299 motherboard (not existing yet) and a good 6 or 8 core CPU. A large monster case to put all these in. And a much more powerful system power supply.
The build process was not that smooth because of the small Manta case, but definitely worth it, Manta is sexy. A lesson is learned that never make the screws too tight till the last assemble step for everything. Whole machine took roughly about 4 hours. My lovely wife helped build GPU watercool for me:) The tubing for GPU watercool is the most headache part, I offered labor, of course.