Advanced Micro Devices Inc (AMD)

[Elmer Phud]

Ten -

First of all, data corruption really shouldn't happen all that often. Maybe once per week, and that's if you add up all of the servers in a cluster. That's why bandwidth is irrelevant when it comes to retransmitting data. The only thing with retransmission vs. error correction on-the-fly is that the transmitter must keep a sizable FIFO buffer of all its transmitted data until the receiver can send back a "CRC OK" packet.

The bit error rate may be higher for highspeed differential signaling and as I pointed out, the CRC is attached to every packet, so some bandwidth is always consumed. ECC is in parallel and consumes no bandwidth.

About the receiver sending back a "CRC OK" packet, doesn't this add additional delays to snoops where a remote processor provides the data and then must wait for the CRC OK response over a couple of hops? What about snarfs by other processors? Must the processor providing the data wait for a "snarf OK"? (only half serious here).