Companies building future networks have the challenge of addressing the ever increasing and diverse data demands at an affordable cost. Two new paradigms that has evolved as a result are:

  • Processing is moving closer to the data source rather than data moving to centralized compute resource for processing
  • Software running in the cloud executing functions/tasks rather than on specialized hardware or local virtual machine

These factors are shaping the development of today’s networks for the following reasons: strict latency requirements of emerging applications, sheer volume of data, the desire to optimize scarce networking resources and the need to connect billions of devices. As we look to the next decade, network infrastructure will become even more pervasive and integrated with every aspect of our everyday lives with the advent of technologies such as 5G, cloud, data center, to name a few. On one hand, 5G networks continue the paradigm of previous cellular standards in terms of driving bandwidth, but they also extend it to many more devices and usage models. Conversely, the rapid explosion in data-center network traffic driven by software-defined networking (SDN), Open vSwitch (OVS), and Network Functions Virtualization (NFV) demands a new class of evolving accelerator cards with continually greater offload capabilities.

The Speedster7t FPGA family is optimized for high-bandwidth workloads and eliminates the performance bottlenecks associated with traditional FPGAs. Built on TSMC’s 7nm FinFET process, Speedster7t FPGAs feature a revolutionary new 2D network-on-chip (NoC), an array of new machine learning processors (MLPs) optimized for high-bandwidth and artificial intelligence/machine learning (AI/ML) workloads, high-bandwidth GDDR6 interfaces, 400G Ethernet and PCI Express Gen5 ports — all interconnected to deliver ASIC-level performance while retaining the full programmability of FPGAs.

Speedster7t Solution

Many new and unique architecture features in Speedster7t FPGAs enable innovations in different networking technologies including 5G, network accelerators SmartNIC, high throughput packet processing and traffic management, and data-path security. Speedster7t FPGAs offer the highest performance interfaces including PCIe Gen5, fracturable Ethernet MACs up to 400G rates, GDDR6 for the highest bandwidth and lowest cost off chip storage, and DDR4/5 for large offchip bulk storage. Internally, Speedster7t FPGAs have been optimized to support high bandwidth data streams from these high performance interfaces with the new 2D Network on Chip (NOC) that enables efficient data transfers for multi Tbps of bandwidth in, out and across the FPGA fabric. An example of the aforementioned advantages is that Speedster7t FPGAs are the only FPGAs that can Terabit Ethernet Switch (TbE) – a critical network entity for service and internet providers alike.

Speedster FPGAs Uniquely Support Full Bandwidth 400G Ethernet

Until recently, FPGAs have been at the center of Networking applications. In the past, designers would increase FPGA throughput by increasing the size of the buses that move data in the FPGAs. This methodology of increasing bus sizes no longer works. To support a 400G Ethernet stream, the FPGA would need to run a 1,024-bit bus at 724 MHz or a 2,048-bit bus at 642 MHz. Neither of these bus width options and FPGA performance are possible in today’s most advanced FPGAs. However, Speedster7t FPGAs can easily run multiple 400G Ethernet datastreams using the new 2D NoC, which has dedicated circuitry to transfer the 400G data streams to the FPGA as 4 independent 100G data streams that are 256-bits at 506 MHz. Buses of this dimension and performance are easily managed in the Speedster7t fabric.

 

Application Requirement Speedster7t Value
Need for high bandwidth external connectivity
  • Multiple ports of 400G Ethernet
  • Multiple ports of PCIe Gen5
  • Highest Performance SerDes – 112 Gbps
  • XSR SerDes for extremely low interface power connection
Highest memory bandwidth for buffering >1 Tbps Memory hierarchy

  • Up to 16 independent GDDR6 channels at 16 Gbps
  • Up to 4 Tbps of total external bandwidth
Wide and high performance data-path Dataflow fabric optimized for high throughput data transfer and acceleration

  • Up to 20 Tbps of NoC bandwidth for high-speed, wide-data transfers
  • Traditional bit-wise routing plus new optimized 8-bit bus routing
  • Homogeneous FPGA architecture for custom matrix computation
  • Support for arbitrary number formats, routing structures and memory hierarchy
Substantial logic capacity to implement complex, custom algorithms Scalable family of devices up to 2.6M 6LUTs of logic
Higher precision data formats for edge training efficiency
  • FP16, FP24
  • bfloat16
  • Only FPGA to support Block float 16
 FPGA fabric for Compute
  • Highly flexible compute for advanced applications such as beamforming
  • High data bandwidth architecture to support massive compute
  • Parallel packet processing engines to support compute intensive functions such as DPI
Substantial internal memory for queuing pointers, internal cache and scratchpad RAM Up to 300 Mbits of internal SRAM

 

Access Transport (incl. 5G) 5G Fronthaul Network Acceletor (SmartNIC)  Network Test Equipment Security Appliance TbE Switch
Highest Performance SerDes
112G multi-Standard SR/MR/LR PHY Yes Yes Yes Yes Yes Yes Yes
Ultra-Short Reach (USR), eXtra-Short reach (XSR) Yes Yes Yes Yes Yes Yes (Silicon Photonics)
Most Advanced Interface IP
400G Ethernet – 16 lanes each running up to 100G Yes
SyncE, IEEE1588, Time Sensitive Networking Yes Yes Yes Yes Yes Yes
PCIe Gen5, 512 Gbps per port Yes Yes Yes Yes
GDDR6 – 4 Tbps of memory bandwidth Yes Yes Yes Yes Yes
DDR4 – up to 3,200 MHz, 3DS stacked memory Yes Yes Yes Yes Yes Yes Yes
DDR5 – up to 4,400 MHz Yes Yes
Application Specific Interface (e.g. CPRI) Yes Yes Yes Yes Yes Yes Yes
Terabit Speed Routing
NoC Yes Yes Yes Yes Yes Yes Yes
Bus Routing Yes
Fully flexibility bit wise routing
High Throughput Processing
Packet Processor/Traffic Manager Yes Yes Yes Yes Yes Yes Yes
Datapath crypto Yes Yes Yes Yes Yes
MLP Yes Yes Yes Yes
Fine grain hardware reprogrammability (examples listed) P4 – programmable pipeline P4 – programmable pipeline Custom functions Network diagnostics Network diagnostics Application specific cryptography P4 – programmable pipeline