Sign up for our newsletter and get the latest HPC news and analysis.
Send me information from insideHPC:


GigaIO Rolls Out Industry’s Highest Performance Top of Rack Switch and Network

Today GigaIO announced the launch of FabreX Gen4. As the industry’s first networking platform to be fully compatible with Peripheral Component Interconnect Express (PCIe) 4.0, FabreX Gen4 doubles the network bandwidth to 512 gigabits per second at full duplex.

FabreX Gen4 is a major step forward for our customers in driving industry leading performance and cost effectiveness through better system utilization,” says Alan Benjamin, CEO of GigaIO. “As the industry’s first in-memory network, GigaIO’s flagship FabreX platform features direct memory access by an individual server to system memories of all other servers in the cluster fabric. Now, with FabreX Gen4, users will benefit from PCIe 4.0’s increased bandwidth while retaining leading sub-microsecond network latency.”

FabreX is a hyper-performance network that delivers improved resource utilization and a unified, software-driven composable infrastructure, reducing the total cost of ownership. FabreX enables direct communication between multiple servers, natively running Message Passing Interface (MPI) applications, containers and hypervisors. Application servers can utilize accelerators and storage located in other servers without going through their respective central processing unit (CPU) and operating system kernel, reducing the demands on the servers’ CPUs, decreasing the latency in the computing cluster and minimizing overall system costs.

Introducing the New Generation of FabreX Technology

FabreX Gen4 extends the capabilities of PCIe 4.0 for Artificial Intelligence, Big Data analytics and high-performance computing. By delivering near linear accelerator and storage performance gains that scale beyond a server’s physical limitations, users are finally able to employ real heterogeneous computing, leveraging CPUs, GPUs, ASICs and FPGAs located in multiple servers for distributed applications.

FabreX simplifies composing computing clusters on-the-fly, driving the performance and dynamic composability of data centers and high-performance computing environments for a completely customized configuration. Additionally, the fabric creates a re-customized configuration without re-cabling. Connections between compute, storage and application accelerator resources are implemented with the rugged, packetized communication protocol of industry standard PCIe 4.0. Moreover, the connections are backward-compatible with existing PCIe 3.0 servers, solid state drives, GPUs and accelerator cards, protecting users’ existing investments. Users can also mix PCIe 3.0 and 4.0 devices in a single network on FabreX and add PCIe 4.0 devices when they become available.

With the introduction of FabreX Gen4, the inherent advantages of the FabreX architecture become more apparent,” continues Alan Benjamin. “We now offer both PCIe 3.0 and 4.0 hardware, and all of our host and switch software run on both platforms. This allows our customers to easily tune their clusters to exactly the performance and flexibility characteristics they want. Whether it’s PCIe 3.0 in the composition and leaf layers, and 4.0 in the spine, or 4.0 throughout, customers can now dial up the performance and utilization of their systems to drive a significantly lower total cost of ownership.”

FabreX supports native non-volatile memory express over fabrics (NVMe-oF), enabling expedited access between cluster compute and storage sub-systems for bandwidth and latency performance that exceeds legacy networking technologies. These technologies are critical for supporting today’s demanding heterogeneous application workloads. FabreX is managed using the industry standard Redfish® APIs, enabling maximum compatibility with existing orchestration and management workflow tools.

As data becomes more demanding and complex, the computing industry is struggling to find a solution that can drive utilization while keeping operating costs down,” says Dan Olds, principal researcher at Gabriel Consulting Group. “GigaIO solves this problem – FabreX is the only solution on the market that can take expensive units, currently in siloes, disaggregate them and make them available so people can share open resources cost effectively.”

In this video from SC19, Alan Benjamin from GigaIO describes how the company’s FabreX Architecture adheres to industry standard PCI Express (PCIe) technology, integrating computing, storage and input/output (IO) communication into a single-system cluster fabric for flawless server-to-server communication and true cluster scale networking.

Availability

FabreX Gen4 includes all the components necessary to build a complete network, including a top-of-rack switch, Hub Class switch, server adapter cards, cables and server-based drivers. FabreX Gen4 will be available in the second quarter of 2020.

Sign up for our insideHPC Newsletter

Resource Links: