Any info on the SATA-IDE adapter? Does it use a Marvell chip or JMicron chip on it, or something else?
There is absolutely no excuse for these to have a compatibility issue with an i815 chipset, but it is what it is. You could try the opposite brand as that helps sometimes.

After waiting almost two weeks for some "new" IDE cables I solved the problem. For anyone stumbling accros this post:

The old cables had only 40 (thicker) wires with 3 black connectors which don't support DMA.
The new ones have 80 (fine) wires with a blue, a gray and a black connector.

Not true. 40 conductor cables support ultra DMA, just not ATA100 or better.

They can totally do ATA33, and sometimes, if they are short enough, ATA66.

For ATA100, ATA133, or ATA166, you need an 80 conductor cable.

Thanks for the clarification, didn't know that. But then I don't know why it works with the 80 and not with the 40 conductor cable. Position on the cable and jumper settings of the devices were the same.

Softwate limitaton? The PC/BIOS can detect an 80 conductor ribbon cable, as far as I know.
Because one of the PATA wires is grounded, I think.

That's why 80 conductor cables aren’t so usable for 1:1 connections.:
40 conductor cables are more universal here, they can be used in non-PATA applications, too.

Except for the tap in the middle with another IDE connector, which then simply has to be left unused.

Sounds like the old 40 conductor cable ran ok at 16 mhz pio4, but when you enabled udma on the bridged sata device, it switched to 33mhz and was not electrcally adaquate. By “not adequate” I mean it wasn’t clearly transmitting the signals and so there was a large amount of retry requests to the ide devices.

Why? It could be a borderline cable that wasn’t quite up to spec any more, it could be the multiple devices, it could be the pata/sata bridge was doing something odd and it tried to go faster than 33 mhz.

douglar wrote on 2025-07-26, 10:52:

so there was a large amount of retry requests to the ide devices.

This definitely makes sense. People should try (vintage) Linux in cases like this when testing, since it logs all such errors to dmesg. Not to mention hdparm allowing finer control of the parameters for testing.