here's what we have: the alpha is connected to a 3com 3900. the 3com 
believes the alpha is running full duplex:

          linkStatus           macAddress          autoNegMode                 
              enabled    00-80-3e-5c-05-b9         enable                 

         autoNegState          reqPortMode       actualPortMode                 
            completed              100full       100full                 

On the netbsd side, I have this: 
fxp0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 1500
        address: 00:a0:c9:e4:63:95
        media: 100baseTX status: active
        inet 130.33.113.34 netmask 0xfffff800 broadcast 130.33.119.255

note that there's no indication of full duplex. I only know how FreeBSD
prints status, which is to show the duplex mode in the media type. Maybe
the newest netbsd ifconfig shows this too, but: the newest ifconfig won't
compile on this system (sup in progress at this end, I hope to fix this
soon). BUT: for the kingston cards, the FDX light is lit, leaving me to
believe that at least those cards are running FDX, even though: 

de0: flags=8863<UP,BROADCAST,NOTRAILERS,RUNNING,SIMPLEX,MULTICAST> mtu 1500
        address: 00:c0:f0:31:61:5a
        media: 100baseTX status: active
        inet 130.33.113.32 netmask 0xfffff800 broadcast 130.33.119.255

again, no full duplex indication from ifconfig, but the FDX light is lit 
on this card. 

going back to the system with the fxp0 in it, the netstat shows: 
Name  Mtu   Network       Address              Ipkts Ierrs    Opkts 
Oerrs  Coll
fxp0  1500  <Link>        00:a0:c9:e4:63:95     5170     0    10665     
1   100

(100 collisions)

after a netperf run, we have:
0  1500  <Link>        00:a0:c9:e4:63:95     5708     0    14567     1   141

so 41 more collisions, for an additional 660 or so Ipkts, and 4000 or so 
Opkts. 

The machine with the de0 shows the following:
Name  Mtu   Network       Address              Ipkts Ierrs    Opkts 
Oerrs  Coll
de0   1500  <Link>        00:c0:f0:31:61:5a   115046   177    90353     
3     0
de0   1500  130.33.112/21 130.33.113.32       115046   177    90353     
3     0

so O(100K) packets, 0 collisions, 177 Ierrs. Things seem to work right, 
just very slowly. 

We had to completely reload the SRM bios on these machines as they came 
with ARC bios. Is there any chance at all that some environment variable 
is set wrong and causing problems? 

Now from the linux side, we have reports from DCG that on this
motherboard, using the linux de4x5x driver, NOT the tulip driver, and on
the latest kernel (2.0.34) incorporating "estabrook's alpha patches" (??),
~ 100 mbits/second is routine. We're trying that now, via reloading linux
on one of the machines. 

As for the machine at the other end, it is in FDX mode. 

most mysterious.

This is the 0829 kernel, compiled from sys.tar.gz at ftp.freebsd.org

For what it's worth we have 144 machines in the same room (non-alpha)
running freebsd and linux and working fine in 100baset full duplex mode. I
think the problem does lie in the alpha side of the house, we just can't
figure out where. We've tried all the obvious things, now we're moving on
... 

ron

Ron Minnich                |"Using Windows NT, which is known to have some 
rminnich@sarnoff.com       | failure modes, on a warship is similar to hoping 
(609)-734-3120             | that luck will be in our favor"- A. Digiorgio
ftp://ftp.sarnoff.com/pub/mnfs/www/docs/cluster.html