NVMe over TCP

Demand for high performance compute and memory is becoming increasingly difficult to scale, as throughput, latency and capacity demands for both are harder to meet. Compute is no longer just CPU; it is heterogeneous computing, which commonly means physically separate hardware compute units. Domain specific compute units have challenges accessing each other’s memory space. NVMe over TCP offers a solution that brings the system components together, allowing it to operate with far more freedom and flexibility for scaling to changing requirements.

NVMe over TCP is a protocol designed to meet high performance requirements and run on standard Ethernet infrastructure. It offers extremely low latency access, to a vast amount of remotely located storage. Introducing routable TCP as the carrier for NMVe means that the physical separation of compute and memory is no longer an obstacle. The constraints on distance, power and cost will, in many cases, allow systems to scale far better with Ethernet based hardware and software, without vendor lock-in to specialist expensive and low volume hardware. The NVMe over TCP protocol has been adopted by Linux and is becoming increasingly popular among other established vendors, particularly within the network storage community.

Moving data around a system consumes time, money and energy; cutting out unnecessary steps reduces the cost of all three. FPGAs can, with the right software and IP, outperform other solutions in terms of throughput, latency, power and cost. Some problems in AI/ML, machine vision, video coding, genetic sequencing, communications and others, can maximise performance in all these variables only by using FPGAs. It is possible to remove a large amount of power hungry data copying in the path from NIC, CPU, DDR memory, PCIe to FPGA. Going directly from memory to FPGA over Ethernet, eliminates excess data duplication.

Chevin Technology provides FPGA IP solutions that reduce the total cost and development time for Ethernet enabled applications. Using Chevin Technology’s IP functions for FPGAs helps developers speed up their development, by making use of existing functions, tools, libraries and examples to reduce risk and development time. This results in a lower overall development cost and time to market.