Data transfer speed is a major bottleneck for critical applications such as analyzing video, modeling complex systems, and training machine learning (ML) algorithms. Although high-speed processors and optical networks operate at terabits per second (Tbps), the network interfaces that transfer data between computers and over networks operate at much slower rates of gigabits per second.
DARPA’s Fast Network Interface Cards (FastNICs) program is aimed at the development, implementation, integration, and validation of novel, network interfaces that can achieve 100-fold speed-ups for distributed training of ML models and other applications. Fast Lanes for Expedited Execution at 10 Terabits (FLEET) is Peraton Labs’ solution for FastNICS. FLEET’s innovative hardware and specialized software leverage industry standard interfaces, reconfigure computing topology in real time, and enable computers to communicate at speeds up to 10 Tbps.
High performance computers (HPC) are typically optimized to handle specific applications. The topology of the interconnects as well as the computational hardware, memory, storage, and the manner of writing software programs are all designed for a specific application. Using new optical network interface cards (O-NICs), which are PCIe devices with multiple photonics links on a card, and programmable optical switching, FLEET adapts the topology in real time to dynamically create a flexible, high performance computer that supports many different applications on the same commodity hardware.
The FLEET planner harnesses this flexibility by automatically configuring available resources for optimal performance tailored to each application. Using specialized software—resource managers, application programming interfaces, and programming tools—developers write applications without having to consider details of the hardware environment. The planner, which leverages the open-source Legion system, reduces cost and saves time by eliminating the need to design, configure, and test a custom topology for the application.
Initial FLEET O-NICs have four photonics links and run at PCIe gen3. By the end of 2023, O-NICs will provide 1 Tbps bidirectional link speed per card using PCIe gen5; in the future, this will be 4Tbps with the emerging gen7 standard. FLEET also works with other networking technologies, such as NVIDIA NVLink for intra-server GPU connectivity.
In addition to speed and reconfigurability, FLEET offers these important benefits:
Learn more at: [email protected].
DARPA’s Fast Network Interface Cards (FastNICs) program is aimed at the development, implementation, integration, and validation of novel, network interfaces that can achieve 100-fold speed-ups for distributed training of ML models and other applications. Fast Lanes for Expedited Execution at 10 Terabits (FLEET) is Peraton Labs’ solution for FastNICS. FLEET’s innovative hardware and specialized software leverage industry standard interfaces, reconfigure computing topology in real time, and enable computers to communicate at speeds up to 10 Tbps.
High performance computers (HPC) are typically optimized to handle specific applications. The topology of the interconnects as well as the computational hardware, memory, storage, and the manner of writing software programs are all designed for a specific application. Using new optical network interface cards (O-NICs), which are PCIe devices with multiple photonics links on a card, and programmable optical switching, FLEET adapts the topology in real time to dynamically create a flexible, high performance computer that supports many different applications on the same commodity hardware.
The FLEET planner harnesses this flexibility by automatically configuring available resources for optimal performance tailored to each application. Using specialized software—resource managers, application programming interfaces, and programming tools—developers write applications without having to consider details of the hardware environment. The planner, which leverages the open-source Legion system, reduces cost and saves time by eliminating the need to design, configure, and test a custom topology for the application.
Initial FLEET O-NICs have four photonics links and run at PCIe gen3. By the end of 2023, O-NICs will provide 1 Tbps bidirectional link speed per card using PCIe gen5; in the future, this will be 4Tbps with the emerging gen7 standard. FLEET also works with other networking technologies, such as NVIDIA NVLink for intra-server GPU connectivity.
In addition to speed and reconfigurability, FLEET offers these important benefits:
- FLEET supports automated, peer-to-peer PCle sharing across the entire datacenter and requires no extra developer work.
- FLEET can use photonic amplification or tunneling to synthesize global-scale, virtual datacenters, supporting remote, direct memory access to computers anywhere in the world.
- FLEET’s synthesized, reconfigurable, virtual datacenters offer significant improvements for disaster recovery by dynamically restoring resources to priority jobs and streaming in-progress results to backup locations.
- FLEET O-NICs contain a powerful field-programmable gate array (FPGA) which can host user code and perform energy-efficient, in-flight data manipulation and transformation.
Learn more at: [email protected].
