The graph is based on approximately 100 calls to the Demon K56Flex enabled Ascends using a mixture of Pace 56k and Hayes 56k modems. No obvious distinction in raw connection speed was noticed between the two. The vast majority of connections using modem defaults were at TX 26400 and RX 46000, but by S91 tweaking I was able to get TX 28800 rates most of the time. A few K56 connections dropped to the ludicrously low data rates of TX 12000 RX 12000 for no obvious reason (perhaps a firmware glitch?). Note that about 30% of calls drop to RX 44000 and a few drop to 42000 and 40000 in an exponential tail. K56Flex never seems to negotiate to higher rates than the initial connection.
V.90 typically connects at 49333 under the same line conditions onto V.90 enabled Ascend hardware and 50667 when connected to 3Com/USR V.90 hardware. The results here are for 100 successful connections. There was no difference between the Pace and Hayes modem initial connect speeds, but the Pace fails to connect about 40% of the time during initial training. The Pace V.90 locks up in a noisy trilling battle with the Ascend (this does not happen on the USR kit). V.90 connections negotiate speeds up and down relative to the initial connection rate and sometimes erroneously get a spurious 56000 maximum RX rate. Measured throughput is generally better than with K56Flex, but is sometimes bursty as the modems are more inclined to step rates with changes in line quality. I have noticed that compressible material gives the expected rates, but that V.90 seems to struggle a bit with incompressible binary data. This may be an illusion though.
I suspect that it may be advantageous to limit the maximum allowed RX data rate to be at most one slot above the usual connection rate reported for your line as this will prevent fall forward delays. TX rates with V.90 remain stubbornly stuck at 26400 or 24000 and do not respond to S91 tweaking at all. I would welcome reports of successful optimisation of V.90 rates using S91 gain adjustment or any other method.