Geoff Tate is CEO and co-founder of Flex Logix, Inc.
FPGA chips have increasingly emerged as a valuable tool in creating a reconfigurable data center: reconfigurable networking adapter cards, reconfigurable networking systems products, and reconfigurable data center projects such as Microsoft’s Project Catapult, Microsoft’s Azure SmartNIC and Amazon EC2 F1.
However, it has recently been proven that embedding this technology on chip can speed performance even more. That is why we are now seeing a host of embedded FPGA companies emerge. This was also the main driver behind Intel paying $16 billion for Altera. Through this strategic acquisition, Intel can first integrate Xeon processors and FPGA chips in multi-chip packages and then eventually integrate FPGA onto the same die as the CPU.
With all this activity, it is clear that embedded FPGA is making its way into the data center in a rapid way, and not soon enough considering the bandwidth demands that are now being placed on them. Every data center manager today wants to build their facility to be more reconfigurable and programmable as opposed to the old PC-era age of hardware that needs to be replaced to keep up with the changing standards and protocols. Embedded FPGA allows this to happen for the first time. This is a win-win for everyone since the data center gets the flexibility that they need and the chip designers can extend the life of their chips and target more markets with those same chips.
Embedded FPGA vs Traditional FPGA
While standard FPGA has delivered many benefits to the data center, it does not solve the performance bottlenecks in chip-level data center usages in the PCIe bus. However, when the FPGA is integrated in a multi-chip package, the bus widths and speeds possible through a interposer are 10x higher than over the PCIe bus. This performance can go up yet again if the FPGA is integrated onto the same die because the buses possible on silicon are much wider and faster than interconnects on PCB boards/connectors.
Also, there are applications such as packet parsers where FPGA is useful in many, distributed small chunks integrated in the networking chip.
What is Embedded FPGA?
Embedded FPGA is now available and silicon has been proven in multiple, popular process nodes. This technology is quickly finding a home in micro-controllers, IoT, deep learning, SoCs/ASICs, wireless base stations as well as in networking/data centers. However, many people still don’t know the difference between this technology and mainstream FPGAs offered by companies such as Xilinx. In the following few paragraphs, we’ll discuss some of the differences.
The basic building block of an FPGA is a LUT, or look-up-table, which when combined with an optional, by-passable flip flop enables programmable logic of any function. The programmable logic is available in a fabric of programmable interconnect which allows any logic block to connect with any other logic block. Verilog or VHDL are used to code FPGA: software maps the code onto the FPGA fabric and produces a bitstream which when loaded in the FPGA causes it to execute the RTL as desired.
Networking/data center applications require 28nm or 16/14nm performance and only three suppliers have embedded FPGA in these advanced processes. The big performance benefit of embedding FPGA is the bandwidth the FPGA can have with the rest of the chip. Even a small embedded FPGA, such as 5K LUTs, can have 1000 inputs and 1000 outputs with data rates of >500MHz in 28nm or more in 16/14nm.
In an FPGA chip, typically there is a mix of programmable logic blocks (logic cells = LUT + flip-flop), DSP/MAC (16×16 or larger multiplier-accumulators, useful for signal processing or deep learning), and Block RAM (a reconfigurable dual-port memory).
With embedded FPGA, there is also more choices on the ratio of logic/MAC/memory. This is very important because in high volume chips, customers don’t want to have functionality they don’t need.
In some cases, chip companies may prefer initially to use an FPGA in a multi-chip module, such as in Intel’s first use of FPGA with Xeon. There are also suppliers that offer support to implement “FPGA Chiplets” for 2.5D multi-chip packaging. Chip-to-chip performance is less than on-chip buses, but much faster than the PCIe bus.
Example Applications of Embedded FPGA in Data Centers
While embedded FPGA is new, we expect volume applications to emerge in the next few years, including:
- Xeon with integrated FPGA
- SmartNICs (a la Microsoft)
- Reconfigurable switches with programmable packet parsers
- Deep learning
Harvard University recently announced it is integrating embedded FPGA into its deep learning chip in 16nm, which is now in fabrication. The reason is deep learning algorithms are evolving rapidly; with embedded FPGA, the algorithms can evolve in real time, not once a year when a new chip tapes out.
Every data center in the world will benefit from being more programmable and reconfigurable in order to keep up with the constantly changing standards and protocols. In the past, data centers typically had to wait for new hardware and chips to meet these new requirements. This is all changing with embedded FPGA and data centers will soon find themselves able to quickly and seamlessly stay ahead of the curve, and ahead of the world’s bandwidth demands.
Opinions expressed in the article above do not necessarily reflect the opinions of Data Center Knowledge and Informa.
Industry Perspectives is a content channel at Data Center Knowledge highlighting thought leadership in the data center arena. See our guidelines and submission process for information on participating. View previously published Industry Perspectives in our Knowledge Library.