partner-offering-banner.png

AI Acceleration Infrastructure   

Integral & Open Systems,Inc

Select

Integral & Open Systems (IOS) designs custom AI Network Interface Cards, SmartNICs, and high-throughput data-plane subsystems on Altera Agilex 7 and Agilex 9 devices for AI cluster, distributed inference, and accelerated database environments. Our designs scale to high-rate (up to 800 GbE class on capable devices) line speeds, with hardware support for tensor pre-processing, in-network reduction, collective operations, and protocol offload — moving the right work off the host and into the fabric where it cuts AI training and inference tail latency.

Engagements cover end-to-end data-plane design: high-speed serial I/O (10/25/100/400 GbE class with paths to higher rates on capable devices, plus PCIe Gen4/5 and CXL), packet processing pipelines, hardware-offloaded RDMA, congestion control, and AI-specific operators including all-reduce, all-gather, and tensor sharding primitives. We also build host driver stacks, kernel-bypass userspace integrations (DP...

Integral & Open Systems (IOS) designs custom AI Network Interface Cards, SmartNICs, and high-throughput data-plane subsystems on Altera Agilex 7 and Agilex 9 devices for AI cluster, distributed inference, and accelerated database environments. Our designs scale to high-rate (up to 800 GbE class on capable devices) line speeds, with hardware support for tensor pre-processing, in-network reduction, collective operations, and protocol offload — moving the right work off the host and into the fabric where it cuts AI training and inference tail latency.

Engagements cover end-to-end data-plane design: high-speed serial I/O (10/25/100/400 GbE class with paths to higher rates on capable devices, plus PCIe Gen4/5 and CXL), packet processing pipelines, hardware-offloaded RDMA, congestion control, and AI-specific operators including all-reduce, all-gather, and tensor sharding primitives. We also build host driver stacks, kernel-bypass userspace integrations (DPDK, AF_XDP, libfabric), and management-plane software for telemetry and provisioning. Designs are validated against customer-defined performance envelopes (line-rate, p99 latency, power) before silicon hand-off.

IOS's value to AI infrastructure customers is the combination of FPGA data-plane engineering with our AI/ML, cloud, and DevSecOps teams — the AI NIC we design lands inside a defensible release pipeline and integrates with the customer's training stack rather than living as a stranded hardware island. The company is CMMC Level 2 self-certified, JCP-certified (DD Form 2345, Cert #0085055), GSA MAS contracted (47QTCA23D00CP), and a DARPA SBIR Phase II prime contractor — a fit for federal AI infrastructure programs as well as commercial AI cluster builders that need defense-grade engineering discipline.

Key Features

  • High-rate Ethernet I/O up to 800 GbE class on capable Agilex 9 devices
  • PCIe Gen4 / Gen5 host interface with CXL support
  • Hardware-offloaded RDMA (RoCEv2 class) with congestion control
  • AI collective operators (all-reduce, all-gather, tensor sharding)
  • Kernel-bypass userspace integration (DPDK, AF_XDP, libfabric)
  • Telemetry and management plane for cluster operations
  • SystemVerilog / UVM verification harness with line-rate traffic generators
  • CMMC L2, JCP-certified delivery for federal AI infrastructure programs
Expand Close
Keyfeatures main image

Offering Brief

Offering Brief

Device Family Agilex™ 9 FPGA Direct RF-Series, Stratix® 10 TX FPGA, Agilex™ 7 FPGA I-Series, Stratix® 10 GX FPGA, Stratix® 10 DX FPGA, Agilex™ 5 FPGA E-Series, Agilex™ 3 FPGA C-Series, Agilex™ 7 FPGA F-Series, Stratix® 10 SX FPGA, Stratix® 10 AX FPGA
Offering Status Production
Demo No
Integrated Testbench No
Evaluation License No
Design Examples Available No
Compliance No
Latest Quartus Version Supported 26.1.0
Development Language C/C++, Verilog, VHDL
Prerequisites Customer-defined performance envelope (line rate, p99 latency, power budget), host interface target (PCIe Gen4/5, CXL), packet processing requirements, and intended integration with customer AI training or serving stack. NDA in place prior to detaile
Languages English
Target Audience AI infrastructure builders, cloud service providers, federal AI program offices, defense primes, intelligence community sponsors, HPC platform builders, and database vendors building hardware-accelerated query offload.
Hands On Lab False

Architecture and performance-budget study (line-rate, p99 latency, power envelope)

RTL data-plane design: packet processing, RDMA, congestion control, tensor operators

Hardware-offloaded RDMA (RoCEv2 class) implementation

AI operators: all-reduce, all-gather, tensor sharding primitives

Host driver development with DPDK / AF_XDP / libfabric integration

Telemetry and management-plane software

SystemVerilog / UVM verification with traffic generators and reference parity checks

Optional: cloud-side orchestration, AI/ML stack integration, DevSecOps release pipeline

Ordering Information

Documentation & Resources

Market Segment and Sub-Segments