You'll certainly get a more professional answer from someone, but in this context, it's not rocket science.

The things are called "dynamic RAM" or DRAM chips, they are dirt cheap on ebay, ALI oder elsewhere and defined by physical package and internal geometry.
The former being "DIP 20" für beetles with 20 feet and the latter most likely 256k x 4, resulting in 1024 kBIT each or 1/8 of a megabyte, thus eight of them resulting in a second MB of memory.

Please check for plausibilty, since I'm hung over, but it should be this:
https://www.silicon-ark.co.uk/aaa1m304p-07-25 … Kx4-dynamic-ram

Also, to check, at least the geometry should be the same as the fixed 8 pieces, so simply google their fine print for a data sheet or merely the same "256 x 4" in the google snippet.

Yes you're right. It looks like it's the same type of memory that is already installed on the card. (barely visible on the picture).
256k x 4bit * 8 chips = 1024k * 8bit = 1Mb
Apparently, the card wants 60ns FPM which is... quite expensive (at least the offers i saw). The 1Mb set would cost me between 20 an 30€... well i bought the card 5€+shipping...
I will keep an eye on dead cards auctions to see if i can scavenge chips from a garbage card...

Thanks a lot !

acl wrote on 2022-03-02, 09:00:

Apparently, the card wants 60ns FPM

It should be a matter of PCI bus frequency which usally is 33 MHz or a certain fraction of the front side bus speed. And, how the card uses the memory. If the memory is too slow, it won't run stable.
(anyone who knows better, please take the time to correct me!)
In 1995, 60 ns was mainstream. So, earlier 30 pin SIMM modules were mostly 70, some late ones 60, then PS2 für Pentiums 60, few 50 ns.
The ones used doesn't mean you'll need them. If you'd find a good offer with 70, I'd give it a try.

acl wrote on 2022-03-02, 09:00:

which is... quite expensive (at least the offers i saw). The 1Mb set would cost me between 20 an 30€... well i bought the card 5€+shipping...
I will keep an eye on dead cards auctions to see if i can scavenge chips from a garbage card...

The qualified "60 ns" probably is what made that offer more expensive. Most don't even provide that info. It's in the numbers on the chips, usually a -6 at the end.
Scavanging is the obvious approach - in my experience, mainstream memory cards of the 90 almost entirely made do with exactly two types of RAM chips, the 128 kB and the 512 kB for up to 4 MB on 2d cards. Consider, if you had 4 MB, you'd also need a big ass CRT to even use that kind of resolution - that being the real cost factor.

But if you buy on purpose: I'm pretty sure I did not pay more than 10 bucks for the set of eight I once got. These are really common, trivial parts. So, I didn't use ALI so for, but for stuff like that, it's probably the better choice - even retail in single digit numbers, a realistic price should be around a buck per piece.
On ebay, if the set says "upgrade for your vintage VGA card" then it probably deserves a premium, for people who go by pin count only. If you understand the geometry like 256 x 4, it's merely a common, bulk industrial component.

I finally found a victim to sacrifice.

I think this could work...

1. De-solder
2. Build a pin adapter for SOJ -> DIP spacing
3. Pullup resistor for the extra address bit
4. Profit (i hope)

At least it was much cheaper than the original discrete chips.

Oh, seriously? I mean, cool if you know how to, but that sounds like a PITA and a gamble.

The offer I linked to, that's out of stock, but the price seems realistic. Nothing in the "industry" section of ebay? Usually you'd find them via the geometry like 256k x 4, the rest by eyeballing...

Yes, i found some, but for some reason, the 60ns are much more expensive than the slowest parts. And i mostly found parts from US with dissuasive shipping fees to Europe. (See at ebay)
I definitely don't want to pay $4.50 x 8 chips + $18 shipping

I got my two ram sticks for 10€. And i consider that i overpaid them, since there are a lot of EDO/SDR sticks for much less. But these were 60ns and the chips were 20 pins and the datasheet looked like it could be compatible. So i will give it a try. This will probably be visually ugly if i succeed.

acl wrote on 2022-05-11, 09:53:

I got my two ram sticks for 10€. And i consider that i overpaid them, since there are a lot of EDO/SDR sticks for much less. But these were 60ns and the chips were 20 pins and the datasheet looked like it could be compatible. So i will give it a try. This will probably be visually ugly if i succeed.

Prices for 72 pin SIMMs are ridiculous anyway, compared to the size. But for around ten bucks, one should get 16 to 32 MB als BIN and most are 60 ns.

But, more importantly, I don't know much about ICs and maybe I'm missing something, but those chips are 1M x 4, not 256k x 4, so how are they supposed to work on that card? Will it just ignore the other 768k?

You seem to be a bit SOL with that particular card, as DIP20 on a PCI card are a bit of an anachronism. Most had DIL40 with 512 KB by then, which are standard in 60 ns.
DIP20 1MBit chips, however, are being sold mostly in 70, 80 or even 100 and 120 ns.
60 ns are rare, so they come at some premium.

You shouldn't expect ebay offers to include the access time in the title. Ebay just isn't that kind of place. There are places that do that kind of thing and include a link to a pdf data sheet in the listing.
- But those either sell you a thousand chips for 500 € or one for 50.

Broader search, I found this:
https://www.ebay.de/itm/222116867106

So, almost 6 € a piece plus shipping from the UK. Still an outrage. I got a set of eight once, pretty sure I didn't pay more than ten for that, albeit 70 ns ones.

I get that you'd like to get the same speed as are on there, on principle. But there are quite a few pics out there with that chip and 70 ns RAM:
https://www.vgamuseum.info/index.php/charts/i … 2-ati-mach64-ct
I think it would be fine with 70 ns, but maybe someone knows more about that.

I'm not 100% confident in my solution but i will post what i found. I coud very possibly be wrong.

The original chips (currently on the card) are Fujitsu 81C4256A-60 256k x 4bit = 128Kb
The ram stick i found are made of HY514404A 1024k x 4bit = 512Kb

Here is a side by side picture of both datasheet. On the pinout section :

Besides the fact that 81C4256A labels its 4bit data pins DQ1 -> DQ4 (HY514404A uses DQ0 -> DQ3). The only difference is an extra address pin.
Pin 5 is not used on 81C4256A (nc -> not connected) while it's used as an extra address line (A9) on HY514404A.

When reading/writing data, row address is loaded first, then column address.

On 81C4256A, we have 9 bit row address and 9 bit column address

• 2^(9 + 9)
• 262144 == 256*1024 == 256k
• So 256k words of 4bit length (256x4) = 128Kb

On HY514404A we have 10 bit row address and 10 bit column address

• 2^(10 + 10)
• 1048576 == 1024*1024 == 1024k
• So 1024k words of 4bit length (1024x4) = 512Kb

So, my theory is that if i fix the A9 address line to zero, the chip will load row and columns on 9 address lines instead of 10. Making the chips virtually the same. (75% of HY514404A capacity is wasted, but i don't mind)

That's a long shot.... But electric signals are compatible. I will probably not burn my card. I must admit that i did not checked the timings... i'm assuming that 60ns EDO should be compatible with another 60ns EDO if the card uses "non aggressive" timings.

Again, i can be 100% wrong. I never swapped DRAM like that.

you don't need adapter. Desolder all the sockets and solder memory directly to the pads using hot air station.

Cheers,

The larger RAMs will not work, CBR refresh is used by the video chipsets and they only do enough refreshes to handle the 128KB chips. With 512KB chips there are lot more rows and they will not get refreshed frequently enough, resulting in image full of noise due to data corruption. You need 128KB chips to have proper result unfortunately, also the chips on that stick on the photo are EDO not FPM which can produce further problems.

Tiido wrote on 2022-05-12, 15:58:

The larger RAMs will not work, CBR refresh is used by the video chipsets and they only do enough refreshes to handle the 128KB chips. With 512KB chips there are lot more rows and they will not get refreshed frequently enough, resulting in image full of noise due to data corruption. You need 128KB chips to have proper result unfortunately, also the chips on that stick on the photo are EDO not FPM which can produce further problems.

I would have been very surprised if they did.

going by the VGA legacy page at least, the chip does work with EDO but I wouldn't bet on that, either:
https://www.vgamuseum.info/index.php/charts/i … 2-ati-mach64-ct

Some old DIP20 beetles, I'm pretty sure I got a set of eight for a tenner, some years ago.

Have those humble, utterly generic semiconductors gone the same way as, say, voodoo or sound cards? Are they getting rare due to freaks like us?

I mean, they got produced by the gazillion for all kinds of uses, not just early 80s VGA cards. Those getting rare, what's next? Friggin DB9 connectors?

If on a budget, what do you think, will 60 ns ones have any benefit over 70 ns?

There should be endless supply of suitable DRAMs in china, these chips were made in huge quantity. There are many many part numbers too, that is the main complication in finding anything, one particular chip may be unavailable but another plentiful.

Mixing EDO and FPM likely won't work, you need same kind of parts that the card already has. 60ns will not do harm, but there is no benefit as the main chip is still doing 70ns accesses that it is confed to.

Tiido wrote on 2022-05-12, 16:52:

There should be endless supply of suitable DRAMs in china, these chips were made in huge quantity. There are many many part numbers too, that is the main complication in finding anything, one particular chip may be unavailable but another plentiful.

That's the thing - I did a fairly broad search around ebay and domestic specialized traders. The best offer I found was 30 € for 10 pieces 70ns of rather in-transparent Chinese origin.
I had assumed that "a buck per piece" was already the retail, single digit amount premium for those. 😉

Tiido wrote on 2022-05-12, 16:52:

60ns will not do harm, but there is no benefit as the main chip is still doing 70ns accesses that it is confed to.

Well, but is it? The card itself has 60 ns so far. Is it documented for that chip or will it always be the very same, with any bios, if there are some cards around with 70ns?
Don't know to much about that - only card I ever upgraded was a vanilla VLB card, where 70ns would do for sure...

You seek places like utsource or contact some actual parts broker. Ebay prices are always very inflated.

Oh, if card is 60ns and you put 70ns parts it will work but there can be occasional snow on the image and especially when temperatures rise.

Tiido wrote on 2022-05-12, 15:58:

The larger RAMs will not work, CBR refresh is used by the video chipsets and they only do enough refreshes to handle the 128KB chips.

You looks much more experienced than me
Is that really an issue ? since i'm fixing the extra row and column address to zero ? (litterally not connecting it to the card, but to vss through a resistor)
I assumed that it would only refresh over the 9 address lines connected, effectively limiting the size of the chip to one quarter :

• Largest chip : 2^(10 + 10) = 1024k words
• VGA chips : 2^(9 + 9) = 256k words
• Larges chip with one "disabled" address line : 2^(10 9 + 10 9 ) = 1024k 256k words

But there should be something i miss.

Tiido wrote on 2022-05-12, 16:52:

Mixing EDO and FPM likely won't work

This was one of my concerns. But since EDO is more recent (and can be seen as something like an "enhanced FPM") i'm really wondering what could go wrong.

Tiido wrote on 2022-05-12, 17:22:

You seek places like utsource

Thanks a lot. utsource was the site i dreamt about for years. And i ordered memory from them.
I found both used and new 81C4256A-60. Same chip actually used on the card. I bought 8 used ones for less than $10. (standard shipping to France was ~$11).
Paypal EUR to USD exchange rate was insulting but anyway...

Even if i will have the good ram chips from utsource, i think i will try to use my de-soldered chips for fun. (If soldering 8 * 20 wire on a perfboard is considered fun...). I made an adapter that is soooo cool that it will be a crime to miss that perfboard + desoldered ram + hot glue mess...

CBR refresh uses the internal refresh counter of the DRAM chip and there are many more rows in the larger chip (and controller doesn't specify which row, it only tells "refresh" and memory chip does rest on its own). All of them must still be refreshed in 16ms (or 32 or 64, depending on partcular chip) period, but the video circuitry refreshes them with assumption there's as many rows as 128KB chip has, and result is most of the rows are not refreshed frequently enough which causes data corruption to happen all over. It doesn't matter where you connect the extra address line, this problem will not change.
I have tried to do this sort of RAM upgrade on my older cards and result is always same, lot of corruption and because of this exact reason, one needs the small chips to have success.

thank you very much for the details. That makes a lot of sense.
I was not familiar with CBR refresh (CAS-BEFORE-RAS).
Only the "RAS" refresh, which is driven externally by a memory controller of some sort to ensure that every row is processed.

Your explanation leads me to a short paper that exposes the differences. Also very interesting.
https://downloads.reactivemicro.com/Electroni … M%20Refresh.pdf

I think i just have to wait for my memory to travel from US. At least, these will work.

Hello

Tiido wrote on 2022-05-12, 17:22:

You seek places like utsource or contact some actual parts broker. Ebay prices are always very inflated.

Just a quick follow-up about the memory expansion.
After you suggestion, i ordered chips from utsource. The shipping was quite fast.
Unfortunately, i was quite busy recently and could not update this post.

Everything works great !
Unlike me, be sure to press the chips in the socket to the maximum. At first, i was hesitant to force. But the memory was not recognized. It worked only after i pushed the chips all the way down, and it requires to force quite a bit against a flat surface.

DOS and Windows both see the 2Mb now, and i'm able to push the resolutions up with better color depth.
Great !

Thank you very much for your help !

Congratulations ~