The Ibex Pro is one supercharged machine that will probably hurt your electric bill.
System76’s fastest Ibex Pro with Ubuntu Server 16.10 packs some crazy horsepower with Intel’s latest 22-core Xeon E5 v4 chips and eight Nvidia Tesla P100 graphics processors.
It’s got the same number of GPUs as Nvidia’s superfast DGX-1, which is being used for deep learning. System76 is targeting the Ibex Pro — which is a rack server — at the same market as the DGX-1. The server has fewer, but newer, CPUs, compared to the DGX-1.
An entry-level Ibex Pro priced at US $9,575 will run Ubuntu Server 16.10, with a six-core Intel Xeon E5-2603v4 chip, 16GB of memory, a Tesla K40 GPU, and 250GB of storage.
A fully loaded system is priced at $114,725, and will come with two 22-core Xeon E5-2699v4 CPUs, 1.5TB of DDR4 DRAM, eight Tesla P100 GPUs, 2TB of NVMe-based SSD storage, and an additional 32TB of SSD storage.
System76 has made its name as an Ubuntu PC maker, but the Ibex Pro is one of its handful of Ubuntu servers. The server is targeted at companies looking to implement machine learning for tasks like speech and image recognition.
High-end GPUs are driving image and speech for companies like Google and Facebook. Those companies have deployed GPUs across servers to analyze images, classify them, and then try to identify objects in a picture. The GPUs provide “learning” capabilities to a server, and the more data it is fed, the stronger a learning model gets.
GPUs are also being for analytics, engineering applications, and scientific research. For example, oil and gas companies value a system with many GPUs for exploration, and the Ibex Pro is much more powerful than workstations with dual-GPU configurations.
The GPUs also make the Ibex Pro great for multimedia or computer vision applications. It can be used for movie production or image editing.
But the whole point of putting so many GPUs together is for machine learning, said Ryan Sipes, a community manager at System76.
The importance of GPUs is growing, and even smaller companies want a piece of machine learning, Sipes said.
But machine-learning — or Nvidia’s GPUs — aren’t cheap, explaining why the Ibex Pro costs so much.
The Ibex Pro ships with Ubuntu; System76 has tested the OS to work with the GPU and server components. System76 can ship the server with other Linux distributions like Red Hat, but the customer will have to work with the company to tune all the components to work with the hardware.
The Ibex Pro does not take advantage of the superfast NVLink interconnect in the Tesla P100 GPUs. That feature, for now, is reserved only for the DGX-1 and Power8 servers. Nvidia’s NVLink is a significantly faster interconnect than industry standard PCI-Express 3.0, which is used in the Ibex Pro.
Server makers are expected to offer servers with Nvidia’s Tesla P100 later this year.
The Ibex Pro server will ship in 62 countries worldwide.