What mainboard do you use?
Depending on your chipset and cache config more RAM may mean less performance because it can't all be cached.

Of course, that's sounds reasonable

I will put a POST image in here.

Looking at https://www.elhvb.com/mobokive/archive/intel/ … TX_67703201.pdf then only 64MB is cacheable 🙁

red-ray wrote on 2022-02-27, 13:37:

Looking at https://www.elhvb.com/mobokive/archive/intel/ … TX_67703201.pdf then only 64MB is cacheable 🙁

So, if I understand it correctly, more than 64MB is utterly useless with gaming.
So what's the smart move here? 1x64MB or 2x32MB?

Yes, your chipset (i430TX) is unable to cache more than 64MB RAM. If you have a single 64MB DIMM, plug that in and remove the rest and you should see a healthy boost in performance. I don't think there's any further performance to be gained by using a single stick versus two sticks.

AFAIK there is no memory interleaving on those S7 Intel chipsets, so one stick is fine. However, if you're determined to extract maximum performance, than you should probably experiment with memory timings in your BIOS? This is where having spare sticks could be useful as different memory reacts differently.

UPD: Oh wait, you have an Intel board. Don't think they have any memory tweaking options in BIOS.

BTW, I think I've read the thing about having more than 64MB RAM hurting performance on 430FX/VX/TX chipsets a long time ago, but I still don't fully understand how it works.

Doesn't the system need to actually use more than 64MB of memory in order for it to hit non-cached regions? Or is memory allocation not that linear, and you can actually use non-cached regions with any operation?

Well if that is the case, I just ordered two 64MB PC66 CL2 memory banks and see what the system will do with it if one is plugged in.
Currently the banks I have are CL3, so I assume there would be some performance improvement.

Not unless you can tweak those timings in the BIOS (doubtful with an Intel or OEM board).

Isn't it that the mainboard autmatically takes over the CL of the memory default settings?
It is for sure not possible to tweak memory timings on this mainboard, or anything at all for that matter.

Aebtdom wrote on 2022-02-27, 13:48:

red-ray wrote on 2022-02-27, 13:37:

Looking at https://www.elhvb.com/mobokive/archive/intel/ … TX_67703201.pdf then only 64MB is cacheable 🙁

So, if I understand it correctly, more than 64MB is utterly useless with gaming.
So what's the smart move here? 1x64MB or 2x32MB?

There's absolutely no advantage to be had from 2 DIMMs instead of one on a single-channel, non-interleaving memory controller. If anything the heavier load on the bus will limit clocking/timing options, although that's hardly an issue in this case as you can't tweak anything anyway.

As for the "64MB is utterly useless", consider that even uncached RAM has 1/1000 the latency of hard disk accesses, so what is ideal depends on your operating system and how much memory you are actually using:

DOS or Win9x, using less than 64MB: max 64MB perfoms best, without exceptions.
DOS or Win9x, using more than 64MB: the smallest amount of RAM that covers what you actually use performs best (i.e. if you need 75MB, go for 64+16=80MB)
WinNT/2K/XP or Linux: performance hit from uncached RAM is limited as active stuff is handled in the cached part. Probably more RAM won't hurt - but if you're using <64MB, you might as well stay under it.

Post-2000 using a 430TX for daily driver duties, I found going over 64MB was far preferable to the disk thrash as browsers etc continued to bloat. Keeping it light and tight, most games actually worth playing on a 430TX based system aren't gonna want more than 64.

jheronimus wrote on 2022-02-27, 15:25:

Doesn't the system need to actually use more than 64MB of memory in order for it to hit non-cached regions? Or is memory allocation not that linear, and you can actually use non-cached regions with any operation?

IIRC, Windows allocatesnot a fixed amount, but a certain percentage of RAM as HDD buffer - so if you upgrade to more, chances might increase that the actual program gets pushed to uncached area, the cached area being fixed.

Just a guess - last time it was relevant, I coundn't afford more than 64 MB of RAM.

Add k6-3 and use as much memory as you want

Aebtdom wrote on 2022-02-27, 15:37:

Well if that is the case, I just ordered two 64MB PC66 CL2 memory banks and see what the system will do with it if one is plugged in.
Currently the banks I have are CL3, so I assume there would be some performance improvement.

pc100 CL3 = PC66 CL2
SDRAMs have little eprom on them with a table of minimum acceptable timings per frequency, you can see it using CPUZ (SPD tab) https://www.cpuid.com/news/66-cpu-z-vintage-edition.html

dionb wrote on 2022-02-27, 18:32:

WinNT/2K/XP or Linux: performance hit from uncached RAM is limited as active stuff is handled in the cached part. Probably more RAM won't hurt - but if you're using <64MB, you might as well stay under it.

I have this vague memory of Linux allocating ram from the top down

rasz_pl wrote on 2022-02-27, 23:43:

[...]

I have this vague memory of Linux allocating ram from the top down

Now I'm trying to find something concrete about it and failing, at least wrt Linux.

For Win9x it's simple: top-down, so most accessed OS functionality will be uncached. There's an interesting tool that apparently can help avoid this (by eating up all the uncached RAM right at the beginning of init, and releasing it after Windows' GUI stuff has been loaded into cached RAM), but haven't tested it yet (and particularly in case of Win95 I'm very suspicious of interfering with already very fragile OS processes):
http://www.redline.ru/~ipl/win2cache_eng.htm

WinNT is more complicated, I read reports of it filling bottom-up, but also that it's just top-down too, but that the file caching in RAM is so effective it negates the hit of OS stuff in uncached areas.

Tbh I'm beginnng to itch to test this. A nice system with i430TX and 64/128MB with DOS, Win98SE (and a second install with that win2cache) and Win2k installs (maybe some old Linux for good measure too) and then Doom/Quake/UT Timedemo.

Well, today I received my 64MB CL2 module and ran a test with 3dm99 max.
The result was better then I was expecting.

1x 128mb pc100 CL3 gave 1291 3dmarks and 1554 CPU

2x128mb pc100 CL3 gave 1150 3dmarks and 1507 CPU

1x64mb pc66 CL2 gave a whoppin 1369 3dmarks and 1645 CPU

*edit* found cpu marks with 2x 128mb and corrected the score.

I will try with an extra 64MB tomorrow, so I know if it is the 64mb module CL2 or that 128mb is just too much.

sweet spot would be run it at 100mhz FSB instead of 66mhz. The CPU will run at 250mhz. Your board looks like it can't do that though.

I know, that's why I used 66mhz in the title. And 100mhz isn't exactly where the chip was designed for 😉