4U enterprise Device That Aggregates up to 16 FPGA Boards
The Ultimate Platform for FPGA Networking and Acceleration features
- 16 Hot-swap FPGA boards in 4U device
- 10 ns board-to-board interconnect
- Dynamically changeable board interconnect topology
- 4 Tbps of non-blocking bandwidth
- AMD EPYC 2nd generation CPU
- Support for any PCle form factor board: Full Height / Full length / Double slot
As other LDA devices, NF3 is powered by proprietary Direct SerDes Access (DSDA™) technology which allows FPGA boards to connect directly without a “CPU-in-the-middle”, forming 3D Torus, grids, stars, clusters and naturally supporting vertical and horizontal scaling.
Two independent switching fabrics: standard PCI Express (1) and a Layer 1 interconnect (2). (1) provides up to x8 connectivity to each board and can be dynamically disabled or switched between x2, x4 and x8 modes. (2) links up to 320 ports, supports any data rate up to 12.5 Gbps and provides 4 Tbps of non-blocking bandwidth between 16 cards.
NF3's dynamically configurable Layer 1 interconnect allows selecting various topologies best suited for the use case, such as 3D Torus, daisy chain, star, point-to-point, mirroring, etc. This flexibility makes NF3 a perfect platform for HPC tasks like genomics research, FFT, AI inference, cryptography, blockchain acceleration, image recognition, complex simulations, etc.
Each board sits in its own hot-swap bay.
8 in the front and 8 in the back. All are compatible with any off-the-shelf FPGA board from any manufacturer (PCIe form factor: full-height, full-length and double-slot).
All bays are synchronized through a common reference clock, jitter-cleaned per each bay.
Once inserted, each bay run-time connects to the switching system through twenty 12.5 Gbps lanes. 16 go to the PCIe connector, the other 4 go to the SFP+ or QSFP ports on its front panel allowing direct external access to the Layer 1 interconnect fabric. Data delivered from external port can be broadcasted to all 16 boards right away.
NF3 is equipped with latest generation dual socket AMD EPYC Rome extended ATX motherboard capable of sustaining data rates required for all 16 FPGA boards to work at full load. The motherboard is located in a standalone compartment with separate cooling, unaffected by the heat generated by FPGA boards.
NF3 is a versatile hi-rel system that supports 2N redundancy and automatic failover. It uses four 2.2 KW hot-swap redundant power supplies delivering up to 225 Watts of independent power into each bay. Special monitoring system handles abnormal power consumption scenarios ensuring several levels of power surge protection.
NF3 platform is a perfect solution for FPGA-based acceleration applications:
- High-throughput processing (terabytes of volatile and non-volatile storage)
- Low-latency deep learning inference
- ASIC prototyping
- Video broadcasting
- Protection from cyber threats (hardware-level kill switch)
- Database acceleration
- Lossless packet capture
- Image recognition
