How FPGA technology is evolving to meet new mid-range system requirements
Multiple trends are sending FPGAs down two distinct development paths. On one path, FPGAs are being optimized primarily to accelerate data center workloads. The data center focus is the next holy grail that the larger vendors are laser-focused on. On another development path, there are the traditional FPGA markets of networking, cellular infrastructure, defense, commercial aviation, industry 4.0 and medical. In these markets many engineers feel they are being abandoned. Their development challenges are quite different than the data center focus that the large vendors are focusing on. Here, designers face an increasingly difficult balancing act as they try to achieve a combination of low power and cost without sacrificing performance and security. Navigating this balancing act requires looking at FPGAs in a new way, using new process technology choices, fabric designs, transceiver strategies and built-in security measures. This has led to a new class of mid-range FPGAs that deliver new capabilities for traditional FPGA developers to leverage.
New Process Technology Choices
One way to reduce power while optimizing the cost of mid-range FPGAs is through the use of new process technologies. For example, using Silicon-Oxide-Nitride-Silicon (SONOS) non-volatile (NV) technology on a 28nm technology node provides a lower power advantage as compared to both SRAM-based FPGAs at the same or even smaller nodes. Previous-generation non-volatile FPGAs using 65nm-and-older floating gate NV technology are more expensive than SONOS. Whereas floating gate technology requires 17.5 V to program using large charge pumps that consume a substantial die area, SONOS technology requires only 7.5 V for programming, so charge pumps can be smaller. This technology enables a smaller die size and contributes to a more cost-effective device.
SONOS technology delivers these benefits by using a single poly transistor stack with a non-conductive Nitride dielectric layer (silicon-nitride, Si3N4) as the charge storage element (see Figure 1). Using this approach, only a very small amount of charge will be lost in proximity to any defect that may exist in the bottom oxide. Because the stored charge is non-mobile in the insulating Nitride layer, most of the stored charge remains where it is, intact. A thinner bottom oxide can be used compared to the floating gate technology, and it can be programmed with lower programming voltages (~7.5 V) and smaller charge pumps. Fewer transistors are required with SONOS than with an SRAM memory element.
Figure 1: SONOS technology. (Source: Microsemi)
SONOS technology improves reliability thanks to its use of a push-pull cell containing an N-channel and a P-channel NV device. The NV devices are not in the data-speed path and are only used to control a standard transistor used as the data-path switch. This provides a large functional advantage because any variation in the NV device threshold voltage (Vt) does not change the switch conductance. The way the devices interact acts as a built-in quasi redundancy, preventing performance degradation over the life of the product.
Power consumption is also reduced. First, the SONOS NV FPGA configuration cell enables two different programmable “configuration” states that control the FPGA data signal path, switching it off and on in a way that optimizes the switch device to provide much lower leakage than a standard transistor. Second, SONOS technology can put a device into a state that turns the supply voltage off to the configuration memories in the FPGA logic block while saving the user’s state in low-power latches. This lowers standby power by approximately two-thirds.
There are two other important SONOS benefits. The first is “instant on” capabilities: because the FPGA logic configuration cell retains its state after power-down, there is no need to reload the FPGA design code when power is returned, and no need for an external boot PROM. Second, unlike the configuration memory in SRAM-based FPGAs that can flip state due to neutron hits, a SONOS device’s FPGA logic configuration is SEU-immune. The SONOS NV charge is stored in the nitride dielectric, which is not susceptible to charge loss from neutron hits.
New Fabric Designs
Another way to improve mid-range FPGA performance is through changes to the programmable logic fabric. This enables devices to meet mainstream performance requirements while consuming one-tenth the static power of competing SRAM FPGAs, and half the total power.
Power and performance trade-offs are involved. As an example, 6-input LUTs can provide some speed benefits, but 4-input LUTs are the better choice for a power- and cost-optimized FPGA in a modern process technology. Meanwhile, as process technology has progressed from 65nm to 28nm and beyond, the delay of wiring has come to dominate logic delay, due to poor scaling of metal wire and via resistance. Widening the wires adds to the die area and cost. So, with each succeeding generation of process technology, inter-cluster delay becomes a significant contributor to the critical path, and the speed advantage of 6-input LUTs diminishes. Ensuring rapid direct connections between nearby LUTs can reduce intra-cluster delay, especially in conjunction with advanced synthesis and placement algorithms. Certain logic functions (such as MUX trees) greatly benefit from the direct connections.
For best results, an FPGA family’s power-performance tradeoffs should be carefully optimized for a core logic supply voltage that is somewhat less than the nominal voltage for the process on which it is manufactured. In the case of 28nm SONOS devices this means optimizing the family for a 1.0V core logic supply voltage, with the option to use the full 1.05 V supply when extra speed is required.
The final piece of the FPGA fabric is the math block, which should support 18-bit multiply-accumulate operations. Power savings are realized through the provision of a pre-adder with a full 19-bit result and an input value cascade chain, and by ensuring that the math block supports reduced precision 9-bit operations, including 9 × 9 dot-product mode. The latter is ideal for use in image processing and convolutional neural networks (CNNs).