most likely they are rebranded slower parts

Do your chips have a marking on the bottom?

Most CCChinese srams are shit, plain and simple. BUT, if you're patient you can start selecting the good ones... out of a gazillion dung beetles. There are, of course, exceptions - some users have reported very good results with them BUT the problem is you never know from which batch yours came.
I have bought over 200("farewell money") and the quality went down quite a bit.
You're way better served buying original ones and speed selecting them, albeit the 512ers may be even more sparse then the 256ers.

Also consider that even among officially produced chips there may be a huge fluctuation of component quality.

Attached is a picture of an Alliance-labeled chip that made it through 50mhz as a tag. It was the only one out of the bunch.

Yes, they have 8076ACVD and JV530111L and TAIWAN. Sorry for the quality of the photo, my mobile phone is not good.

I have tried to select the four best with Memtest. You have 126 possible combinations of 4 out of 9 chips if you do not care about the order (though I think the order also affect performance), too many to test them all. I have tried to find a relation between the errors thrown by Memtest and the chip that originated them with no success.

Yes it's not advisable to go the C(n, k) route. I do also suspect that even the order in such "k set" matters but that is just a guess.
I would rather run a first pass selection process whereas you use 8(4 in single bank) srams you know that work and the sram to be tested only as a tag(cache mode: write back). This is what I did for the tests.
If the remaining/selected srams play nice with each other is not guaranteed but may be likely.
Again, I would "advice" you to buying older, proper 512ers and rather select from them. The only official srams I had that were sub-par were, extremely cool looking, Alliance. No guarantees though.

I am not convinced that your issues are related to the SRAM being China-made or reproduction parts. You are using single-banked cache, which is well known for not working well at 40 MHz with the tightest memory timings. For this reason, I always recommend using double-banked cache motherboards (9 SRAM chips instead of 5). Even 15 ns SRAM chips at 1024K in the double-banked configuration "should" work at 2-1-1-1 at 40 MHz with an AM5x86 at 160 MHz. Compounding your problem is that this ALi chipset particularly sensitive with cache speeds. I've noticed the difference between cache and RAM to be less than on other, similar motherboards (SiS 496, UMC 8881). For single-banked ALi M1487 boards, I wouldn't bother going above 256K because you'll need to slow the timings down. If you want 64MB of RAM, set you 256 L2 cache to write-through mode. If write-through mode is not available, then use MODBIN to unhide this feature.

amadeus777999 wrote:

I would rather run a first pass selection process whereas you use 8(4 in single bank) srams you know that work and the sram to be tested only as a tag(cache mode: write back). This is what I did for the tests.
If the remaining/selected srams play nice with each other is not guaranteed but may be likely.

Do you mean testing the 9 chips one by one in the tag socket along with 4 different chips in the data sockets that I know they work at 40 MHz & 2-1-1-1? Would that work? I would need a board with a 32-pin tag socket. And if the board requires a 32Kx8 or 64Kx8 tag chip, would a 128Kx8 work in their place? This way you would be testing only a quarter or half the size of the chip, wouldn't you?

I have another working PC with this ASC486 board:

I could use it to try the double banked way but I lack the 64Kx4 sram required for tag.

Size Bank 0 Bank 1 TAG

32KB (4) 8K x 8 NONE (1) 16K x 4
64KB (4) 8K x 8 (4) 8K x 8 (1) 16K x 4
128KB (4) 32K x 8 NONE (1) 32K x 8
256KB (4) 32K x 8 (4) 32K x 8 (1) 32K x 8
512KB (4) 128K x 8 NONE (1) 64K x 4
1MB (4) 128K x 8 (4) 128K x 8 (1) 64K x 4

feipoa wrote:

If you want 64MB of RAM, set you 256 L2 cache to write-through mode. If write-through mode is not available, then use MODBIN to unhide this feature.

Write-Through mode is not available and I cannot do that. My BIOS chip is not Flash and I do not have a programmer or an ultraviolet light to erase it.

Bellator wrote:

amadeus777999 wrote:

I would rather run a first pass selection process whereas you use 8(4 in single bank) srams you know that work and the sram to be tested only as a tag(cache mode: write back). This is what I did for the tests.
If the remaining/selected srams play nice with each other is not guaranteed but may be likely.

Do you mean testing the 9 chips one by one in the tag socket along with 4 different chips in the data sockets that I know they work at 40 MHz & 2-1-1-1? Would that work? I would need a board with a 32-pin tag socket. And if the board requires a 32Kx8 or 64Kx8 tag chip, would a 128Kx8 work in their place? This way you would be testing only a quarter or half the size of the chip, wouldn't you?

Dang, yours are 1024s... in my reading turbulence I got the impresssion that you were testing 512ers.
You have to try it - I only do know that using a smaller "tag" than required by the manual, albeit still capable of serving the given cache size, can lock up the system... BUT this is not always the case. See what happens if it is larger.

Well, testing each chip seperately along with a high quality, tested, set is at least a way to make sure the chip is worth something. You would need an SRam tester to do this efficiently which is unfortunately very expensive and rare. I'm talking a tester that can also probe the speed - mere functionality test, which is better than nothing, can be had for cheap.

Bellator wrote:

feipoa wrote:

If you want 64MB of RAM, set you 256 L2 cache to write-through mode. If write-through mode is not available, then use MODBIN to unhide this feature.

Write-Through mode is not available and I cannot do that. My BIOS chip is not Flash and I do not have a programmer or an ultraviolet light to erase it.

I suggest you buy a programmer. Most programmers will also test SRAM. This will save you a lot of frustration. Years ago, before I bought a programmer, I tried to test SRAM chips as you are. I tested about 100 SRAM chips. I thought I found all the bad chips. I eventually bought a programmer and retested the SRAM chips. There were still some bad SRAM chips in my pile of good SRAM chips. Save yourself a headache and buy a programmer. From my experience, there is a 10-15% failure rate for these China-sourced 128kx8 SRAM chips. Testing on the motherboard doesn't always catch the marginal chips. If you get the programmer, you can also write EEPROM chips.

It is highly unlikely that you will get 2-1-1-1 working at 40 MHz with 512 KB of single-banked SRAM cache, especially if you are using this method to test a batch of chips which have a 10% failure rate. If you insist on using your motherboard to test SRAM cache, try 2-1-1-1 with a 33 MHz FSB and 256 KB cache. Use 32 MB of RAM max.

I recently bough 10x 1024's for a system I have, only to realize I hadn't bothered to check the motherboard, so they are still in the package I got them in. So I'll be watching this thread.

I bought some on Alibaba or Aliexpress not too long ago (1-2 yrs). They worked, but there is something about remakes that kinda take away from using them in vintage systems. Its sorta like is a large IC company in China made the K5-PR200 today. How would you feel about using it in a cased system? The SRAM isn't as big of a deal compared to the CPU, but it sits in the back of my mind. I wish the printing was in white rather than a faint yellowy brown.