A purpose-bulit machine to design and train your model to detect and identify “Empty” and “Occupied” Parking spots(a.k.a Deep Learning). At USD 15,000 each, the hardware requirement for someone embarking on a deep learning project also need a deep pocket. After much googling, I came up with a cheaper option. Without a monitor, it bloody expensive for me.
|CPU||Intel – Core i5-6600K 3.5GHz Quad-Core Processor||$228.98 @ OutletPC|
|CPU Cooler||Cooler Master – GeminII M4 58.4 CFM Sleeve Bearing CPU Cooler||$32.89 @ OutletPC|
|Motherboard||Asus – B150I PRO GAMING/WIFI/AURA Mini ITX LGA1151 Motherboard||Purchased For $0.00|
|Memory||Kingston – FURY 16GB (2 x 8GB) DDR4-2400 Memory||$193.78 @ OutletPC|
|Storage||Samsung – 960 EVO 250GB M.2-2280 Solid State Drive||$117.60 @ Amazon|
|Video Card||Gigabyte – GeForce GTX 1070 8GB Mini ITX OC Video Card||$424.98 @ Newegg|
|Case||Cooler Master – Elite 110 Mini ITX Tower Case||$38.89 @ OutletPC|
|Power Supply||Cooler Master – 550W 80+ Bronze Certified Semi-Modular ATX Power Supply||$57.98 @ Newegg|
|Prices include shipping, taxes, rebates, and discounts|
|Total (before mail-in rebates)||$1125.10|
|Generated by PCPartPicker 2017-09-21 15:55 EDT-0400|
To side track, I learn that
CPU and Cooling
The CPU must be compatible with the selected chipset while providing sufficient PCIe support. Consumer CPUs are ideal because the application target is highly fault tolerant and the NVIDIA DIGITS DevBox is acting as a workstation instead of a server. Pairing the CPU with effective cooling is crucial for optimal performance, especially when the GPUs are under peak load. I chose the Intel Core i7-5930K CPU with a Corsair Hydro H60 cooler.
Memory and Storage
RAM is important for handling large DNN files and datasets. The Intel Core i7-5930K CPU can stably support up to 64 GB RAM. An Intel Xeon processor can handle more RAM and allow ECC. However, using the Intel Xeon processor will significantly increase the cost.
Chassis, Thermal, and Acoustic Considerations
Acoustics and heat management are major considerations, especially when deploying the NVIDIA DIGITS DevBox in a normal office environment. A chassis that separates the power supply and disks from the heat generated by the CPU and GPUs is ideal.
The power supply should provide enough power to operate the system components along with some headroom to ensure stable operation. The total dissipated power for all of the system components used in a sample build is between 1,200 and 1,300 watts.
Our sample build uses an EVGA SuperNOVA 1600W P2 power supply that delivers approximately 90% efficiency at 100% load (1,400 watts), ensuring system stability at peak workloads.
Effective deep learning requires multiple GPUs. However, suitable PCIe topology is critical to being able to use those GPUs efficiently. Synchronous Stochastic Gradient Descent (SGD) for deep learning relies on broadcast communication between the GPUs. SGD acceleration needs P2P DMAs to work between devices. This means that all GPUs must be on the same I/O hub with a very fast PCIe switches. Workstation motherboards based on the Intel X99 chipset with a PLX bridge setup can support four PCIe Generation3 x16 cards at either full speed or with minimal drop-off.
The sample build used the ASUS X99-E WS workstation motherboard that supports Intel LGA 2011-v2 CPUs while drawing only 20W.