Fig. 3 shows the FPGA-internal block diagram. The three in¬terfaces, serial AER, parallel AER and USB are drawn in orange. The USB interface, as opposed to the other interfaces, is handling explicitly timestamped addresses. Thus we need monitoring and sequencing units (green) between the two domains. The routing fabric between the three interface blocks allows AEs to be selectively routed between the three interfaces. It also contains simple mapping and filtering units.
The mapping units can add a configurable offset to an AE stream, so that different address spaces can be made non-overlapping. The filtering units allow to select which events are routed to which destination.
All these functional units are interconnected using FIFOs (blue, striped).
ON THE IMPORTANCE OF FLOW-CONTROL
Here we compare the statistics of a Serial AER implementation with flow control and one that simply drops events.
With Flow-Control: Assume we have an event-consumer that can handle event rates up to 125MHz. Thanks to the flow-control scheme, the consumer can block the producer as necessary. In this example we choose a fairly strict requirement that an event is delivered with a delay of more than 1/KS at probability of less than 10-6.
Given a Poisson distributed3 producer, this means that the mean event rate of the producer can be up to 63.7% of the consumer event rate without violating our requirements.
Without Flow-Control: For comparison we assume a consumer that can handle event rates up to 125MHz, but if two or more events arrive within an 8ns (= 1/125MHz) time-slot all except the first
备注2:In this calculation the signal propagation speed for the SATA cables was assumed to be half the speed of light, a rather conservative estimate.
备注3:A Poisson distribution is probably an unsuitable assumption when looking at a longer typical AE sequence. But what is critical is the performance in event bursts. We here take the Poisson distribution for looking at such bursts, typical for address event systems. The mean event rate should then be interpreted as the mean event rate in event bursts.
one are dropped. The probability that an event is dropped shall be no more than 10-3. Under these circumstances a Poisson producer can then have a mean event rate of no more than 4.54% of the consumer rate.
Thus for our practical purposes flow-control gives us about one order of magnitude of actually usable event rate. In an experimental setup it also allows us to handle channel congestion either at the sender or the receiver side.
Further discussion of flow control in address event systems can be found in [9].
V. RESULTS AND CONCLUSION
We have developed an AER interfacing board part of a generic AER communication system suitable for building complex multi-chip AE based systems.