Obsessed with technology?
Subscribe to the latest tech news as well as exciting promotions from us and our partners!
By subscribing, you indicate that you have read & understood the SPH's Privacy Policy and PDPA Statement.
News Categories

Meet the highest density compute accelerator with over 79,800 CUDA cores!

By Vijay Anand - on 19 Mar 2015, 7:52pm

Meet the highest density compute accelerator with over 79,800 CUDA cores!

Have you ever seen so many GPUs in such a compact server rack? Note that what you see here are 16 of NVIDIA's top-of-the-line Tesla K80 accelerator cards - each has a dual GK210 GPU setup - for a grand total of 32 GPUs. Imagine the power draw!

You've probably heard and seen compute intensive servers that boast four or even eight GPUs, but One Stop Systems is the first to cram 16 accelerator boards in a 3U rack form factor. We say "accelerator boards" because this system is capable of accepting boards from AMD, Intel and NVIDIA - either exclusively or even in a mixed array!

That's because this isn't a tradition server rack, but a High Density Compute Accelerator (HDCA) appliance that looks like a server rack. Its all-steel 3U rackmount chassis houses nothing but the accelerator boards, three 3000-watt redundant PSUs, power management system and the PCIe backplane to connect up to 4 host computers/servers through PCIe x16 Gen.3 connections. The HDCA's job is to co-work with your server(s) and instantly add teraflops of compute power - independent of your main server configuration. This allows companies to manage their server computing needs in a modular manner rather than overhauling existing machines - unless they are old in which case you would want to replace them to be able to take advantage of these high speed accelerators.

At the moment from the factory, you get to configure it with NVIDIA's Tesla GPUs or Intel's Xeon Phi coprocessors. If you choose to go with the top-of-the-line configuration, you can outfit the rack with 16 Tesla K80 accelerators which will yield you a net single precision performance throughput of 139.8Tflops with a grand total of 79,872 CUDA cores! Great for data analysis, image processing, weather modeling and other such tasks.

Should your choice of HPC applications differ and benefit the x86 instruction set, you could as easily choose to equip all 16 slots with Intel's Xeon Phi coprocessors. Equipped with a full complement of 16 Xeon Phi 7120P model, you get up to 19.2Tflops of double precision performance, which is nothing to sneeze about.

The 3U rack is configured with four canisters each housing four accelerators. Each canister of cards can be configured to power dedicated host systems, thereby increasing the HDCA appliance's usability and versatility in accelerator configuration.

What's interesting is the configuration of One Stop System's rack that houses 4 discrete canisters that can mount four accelerator cards each. Filling up all the canisters will net you a grand total of 16 accelerators within the 3U rack. All the accelerators in the rack can work in tandem to power up a target host, or you can assign each canister of accelerators to different hosts. With four canisters, that means you can purchase up to 4 PCIe x16 Gen.3 add-on connections to serve four host machines. This is where the design of One Stop System's 3U rack offers diversity within a single physical rack as each canister can be outfitted with AMD, Intel or NVIDIA's accelerators. Even if you opt to stick with a single accelerator brand, each canister can be outfitted with a different complement of cards to meet the exact needs of your host machine.

This is the back view of One Stop System's HDCA appliance. You'll notice the triple stacked 3000-watt redundant PSUs, the powerful exhaust fans and of course the PCIe x16 Gen.3 connectivity to the host machine(s) via the adapter cards.

One Stop System's HDCA appliance starts from US$16,000 - without any accelerator boards. It's certainly expensive, but it's also an equally powerful asset to any organization that's willing to step up compute performance in a big way without too much disruption.

The main hurdle to compatibility is whether the host system's BIOS can handle 16 GPUs. As such it's best to drop One Stop Systems a note to check on those specifics and offer qualification advice before finalizing your acquisition of HDCA appliance. Before we end this news piece, since this appliance is being showcased at NVIDIA's GPU Tech Conference, it goes without saying that it's an excellent accompaniment to step into deep learning research and dabbling in deep neural networks (DNN).

For more configuration options and details, you can check their product pages here and here.