Hi ! >>> The trick is you must have an adapter in there somewhere with >>> a BIOS or a CPU on it, I've forgot which. The IBM memory or >>> the SCSI adapters have these. This part is definitely misleading or misunderstood. - even busmaster adapters with 32-bit addressing width cannot substitute the missing DMA functions for the memory above 16MB - the ROM they supply is for their own function. The BOPT-workaround works even with no other adapter installed in the system than the two memory cards. (BOPT = Bypasses One Problem Temporarily) Also the use of Kingston or Acculogic cards pushes the system over the 16MB-limit. The problem is the 24-bit DMA-chip on Mod. 70 and 80 - since 2^24 = 16.0MB addressing range. This is the range where DMA can be used to transfer data among the memory - if the DMA cannot be used direct addressing (PIO) must be used to transfer data to the locations above the DMA-addressing range. Works as well but is a little slower. A problem on the older models might occur with detection of memory errors. The parity-informations are mainly transported with DMA to detect and handle bit-failures. (Mainly cause an NMI error though - and the system stops with 111 ?????? or such) If the DMA cannot directly access the memory a parity error *might* be undetected. The memory handler invoked with the BOPT-workaround uses the PIO-mode for the error-detection ... the Kingston and Acculogic cards have own parity control integrated in their chipsets. This however has nothing to do with the memory *refresh*, which is directly controlled by the memory subsystem on the planar and on the memory cards. Let's say the system has 8MB on the planar and 16MB on a Kingston card. The planar-8MB are under full control of the boards DMA and parity logic. The 16MB on the Kingston card are on the control of the cards' parity control and the lower 8MB can be accessed directly by the systemboard-DMA - the upper 8MB are used via normal 32-bit direct addressing bytewise. The fastest memory access is that for the planar memory: DMA plus 0 - 1 waitstate make it rather quick. The slowest memory access is that on the range from 16MB - 24MB: bytewise direct-accessing to read from memory and to write to memory plus 1 - 3 wait-states on "channel memory" take some time. Pushing a Mod. 70 / 80 over the 16MB border makes only sense with a real 32-bit operating system, which can handle the different memory addressing models with no problems (like OS/2) - DOS / Windows may have some problems. I ran a Mod. 80-A31 under OS/2 Warp Server with 40MB for quite some time without any problem. It had 8MB on the planar (2 x 4MB), 32MB on an Acculogic card (4 x 8MB), an IBM SCSI controller without Cache /A, an Adaptec AHA-1640 (for tape and CD), an IBM Token Ring 16/4 Adapter /A and an AMS 2-LPT card. Very friendly greetings from Peter in Germany (peterwendt@aol.com)