A watercooled desktop workstation equivalent to 400 CPUs
On Wednesday, Nvidia took the wraps off an interesting AI research product that packs the horsepower of a traditional server rack within the economical confines of a traditional desktop PC chassis.
The company’s new DGX Station is a desktop workstation modeled after the larger DGX-1 AI supercomputer, but is conveniently designed for use at a home, lab, or office desk. The system features one tenth the noise levels of other deep learning setups while delivering performance that is three times faster than some of the fastest Pascal-based workstations currently in use.
3x speedup in deep learning compared with Pascal-based workstations
For instance, researchers who run neural net training, deep learning and data analytics tasks can now compute with up to 480 teraflops of half-precision performance, or a 100-fold increase on large datasets compared with a 20-node traditional workstation.
The system features four Tesla V100 GPUs operating with a total of 2,560 Tensor cores, 20,480 CUDA cores, and 64GB of graphics memory. It is powered by dual 20-core Intel Xeon E5-26989 v4 processors running at 2.2GHz each, 256GB of DDR4 LRDIMM memory, and a 1500W power supply. Other features include three DisplayPort connections for simultaneous output to multiple 4K displays, dual 10-Gigabit LAN, three 1.92TB SSDs in RAID 0 for data storage, and a single 1.92TB SSD for the Ubuntu operating system.
Draws less than 1/20th the power at half the noise level
The chassis measures 20.39 x 10.08 x 25.16 inches (518 x 256 x 639mm) and features a water-cooled design for both CPUs and all four Volta-based V100s, allowing it to run AI supercomputing workloads at less than 35dBa of noise, or whisper-quiet library operation. In comparison, some of the quieter Quadro Pascal-based 4U workstations can operate around 43dBa at 1 meter using top-mounted fans, which is almost double the noise level but significantly lower in performance.
In other words, researchers and professionals in the AI and deep learning fields can now get the computing capacity of 400 CPUs in a workstation that fits on top of a desk and draws less than one twentieth the power of a traditional workstation. Nvidia hopes these metrics will draw developers by the masses, as it outlined earlier this week during its annual GTC conference.
Available now for $69,000
The Nvidia DGX Station based on Volta is obviously a clever, well-designed system that should be marked on every research professional’s wish list. The company will offer enterprise grade support with access to its own deep learning libraries, software upgrades and updates, and priority support for critical issues. It is now available for $69,000 through NPN Accelerated Computing partners, with an order link on Nvidia’s datacenter
