feipoa wrote on 2023-04-04, 11:20:
My statement was only to serve as an agreement - ensure the PCB can be used for 1024K double-banked and 512K double-banked, which it seems like you are planning for.
Yes, 512K with A19 grounded, pulled to 5V (directly or using a 10K resistor) makes an plug-and-play experience to get to 512K without any hardware support. With the builtin resistor, it's actually unplug-and-play 😉 .
feipoa wrote on 2023-04-04, 11:20:
Small word of caution: SRAM module binning may be required to get 1024K working at 2-1-1-1 w/33 MHz and cx5x86. I think user pshipkov experimented with this and found this board to be very SRAM particular. For the quick tests I did with 1024K and 33 MHz, I couldn't get 2-1-1-1 working w/cx5x86.
Now, that's funny. I had no issues with 2-1-1-1 at 40MHz on the Biostar board and RAM WS 0/0 with the 5x86. I did need 1/0 on RAM WS to get stable operation on the HOT-433 with an earlier 8881 revision, though.
This experience is at 256KB cache, but as long as the capacitive load of the chips is approximately the same for 128K x 8 chips, as it is for 32K x 8 chips, stability and timing shouldn't depend on the cache size. On the other hand, there are these notorious IS61C1024N-10 DIP32 chips from China that are definitely fake (in the sense that they were not manufactured by ISSI and specified for 10ns access time), and possibly just way slower than the 32k x 8 chips we typically use for 256K of cache. Nevertheless, the requirement to add a read wait state on the HOT board at FSB40 for the Cyrix, although an AMD 5x86 works perfectly at FSB40 with 0 read wait states confirms that the Cyrix 5x86 is more picky about memory timing than the AMD 5x86.
I didn't spend much time on this because I was targeting 1024K for 66 MHz FSB and 3-2-2-2. I noticed that this board lacks buffers, could this by why 2-1-1-1 was not working well with 1024K? On my other boards, which have the memory buffers, 1024K and 2-1-1-1 was never an issue at 33 MHz.
Indeed, buffers on the memory data lines might help, as they reduce capacitive load on the frontside bus. On the other hand, buffers on the memory data lines also increase memory access time by 3 to 8 ns, depending on the buffer type. Buffers on the memory address lines on the other hand (which are more common than on the memory data lines) are likely unrelated to cache timing, because the row/column-multiplexed memory address lines are completely separate from the frontside bus address lines. Address buffers are more common, because every memory chip connects to every address line, whereas only every memory rank connect once to each data line, so strong drive is way more important on the address lines than on the data lines.
EDIT: Would there be space for SOJ sockets to swap the RAM on your PCB?
No. I'm already using the narrow SOJ variant. It has a 0.3 inch wide body, just as the "skinny DIP" 28-pin and 32-pin cache chips. The distance between neighbouring DIP sockets on the Biostar board is 0.4 inch (which is common), so there is only 0.05 inch space between the edge of the chip case and the border of the socket. This is sufficient for DIP sockets, but as SOJ sockets can only clamp from the outer edge, they are wider. The current layout is created with the fact in mind that the SOJ chips are placed at the same period as the DIP sockets. This will make soldering them an interesting challenge, and is the reason why I ordered a SMD hot-air station and soldering paste in a syringe. I guess I wouldn't be able to solder that board with my home improvement heat gun (it will just blow away the chips), and using an iron might also prove challenging.