L1 cache is internal to the CPU and controlled by it - with no practical limit in case of So7 CPUs (don't know exact amount off top of my head, but >=512MB in any event)
L2 cache is on the motherboard and controlled by a dedicated controller with a given amount of Tag RAM. These limits are completely independent of each other.

There's no point in turning L2 off, it means nothing will be cached at L2 level. Leaving it on will at least cache the first 64MB (or whatever the limit of your board is). The rest will simply be uncached. That will always perform better than turning it all off.

But isn't L2-uncached ram still above hard drive in memory hierarchy, so why having more than 64mb in those chipsets is "bad"? The page says remove ram if you have more than 64mb. If the program requires 128mb ram, isn't L2-cached 64mb + L2-uncached 64mb better than L2-cached 64mb + 64mb swap file ?

The 430HX chipset does 512MB cache if you need that much RAM.

The problem with non-cached RAM is that DOS based OS'es like Win9x uses memory from "the bottom up", thus using the non-cached before cached.

Edit: There's another similar thread right now, where there's a link to some explanation: Re: Cacheable Ram and performance

PlaneVuki wrote on 2023-03-22, 05:31:

But isn't L2-uncached ram still above hard drive in memory hierarchy, so why having more than 64mb in those chipsets is "bad"? The page says remove ram if you have more than 64mb. If the program requires 128mb ram, isn't L2-cached 64mb + L2-uncached 64mb better than L2-cached 64mb + 64mb swap file ?

Using more than 64MB on those systems you cripple your performance, because you can’t control the address space and OS memory management in fact uses the RAM first from the bottom, so it fills the cacheable RAM last. You can get something like 30%-40% performance hit in this case.

So yes, your RAM is naturally still faster than HDD file, but the trade off is that you bring the overall performance down significantly and you can’t mitigate that in any way.

I don’t see 64MB limit a big issue. I have Pentium with TX (64MB limit) and have 98lite installed. It works smoothly and runs everything that is suitable for the system just fine. Heck, back in the day I had 16MB RAM on my Pentium in 1995 and was completely satisfied with it with Windows 95. IMO we easily get distracted with these old systems and try to push them on the level that resembles something we have today, but they are in many ways still limited. Very few home owners (or office workers) had 64MB back in the day.

It is what it is and enjoy the system for what it is. Besides, it is pretty easy to test pentium box with different amounts of memory. In fact, I have planned to do that, I have one unused VX Pentium box sitting around and when I get my other projects done, I jump on it and while tinkering with it, I’ll test different memory configurations out with it and how they affect the performance.

PlaneVuki wrote on 2023-03-22, 05:31:

The page says remove ram if you have more than 64mb. If the program requires 128mb ram

no sane person will ever run anything requiring 128MB of ram on "430FX, 430VX, 430TX" platform 😀
64MB was enough for everything in 1999 https://www.anandtech.com/show/267/6

Similar topic on same subject: Re: Best memory configuration for a Last generation Pentium (Socket 7)
"May 1997:
128MB = $300
Ram 16MB lowend, 32MB standard, 64MB would be only available on the absolute top end.
CPU garbage bin (whole system $499) 120MHz Cyrix. lowend ~133MHz Pentium/K5. Standard 200MHz Pentium. Top 200MHz PPro.

July 1998:
128MB = $150
Ram 32MB lowend, 64MB standard/top.
CPU lowend ~200-233MHz Pentium, 266MHz Celeron, 233MHz Cyrix, 233-300MHz AMD. Standard 300MHz P2. Top 400MHz P2.

tldr: computers with 2.5-3x faster CPU were sold with 64MB of ram because 128MB was too expensive and nothing used that much ram.
Even year later in July 1999 128MB was the TOP ram option, with standard still being 64MB. In PC Mag Jul 1999 the only computers with 256MB ram are $5K workstations with silly silicon graphics cards or $5K, $8.2K and $9K servers."

PlaneVuki wrote on 2023-03-22, 05:31:

But isn't L2-uncached ram still above hard drive in memory hierarchy, so why having more than 64mb in those chipsets is "bad"? The page says remove ram if you have more than 64mb. If the program requires 128mb ram, isn't L2-cached 64mb + L2-uncached 64mb better than L2-cached 64mb + 64mb swap file ?

The hard drive speed when compared to memory is irrelevant, as long as the memory is sufficient for that particular application.
In theory, yes, if you have an app that requires a large amount of memory, then having more memory (even if it's only partially cached) would certainly be better.
If memory is sufficient, then the L2 speed vs memory speed is the limiting factor.

However, as others have pointed out, there is nothing you could do on a Windows 98 PC with a Pentium MMX that would require this amount of memory. If a game can run decently on a Pentium MMX (in terms of processing speed), then for sure it will have very low RAM requirements as well.

Back in 1998, my Pentium MMX 166 MHz came with 16 MB of memory (was using Win95) and I upgraded to a K6-2 500 with 32 MB of memory in the year 2000 (and switched to Win98). There is only one game that I recall which was impossible to run/play with 32 MB of RAM: Nocturne (1999). However, even if you have 128 MB of RAM, Nocturne remains unplayable with a K6-2 500 CPU.

Agreed what others have said. In short, you are more likely to first run into a CPU bottleneck then a RAM limitation with the hardware you described.

PlaneVuki wrote on 2023-03-22, 05:31:

But isn't L2-uncached ram still above hard drive in memory hierarchy, so why having more than 64mb in those chipsets is "bad"? The page says remove ram if you have more than 64mb. If the program requires 128mb ram, isn't L2-cached 64mb + L2-uncached 64mb better than L2-cached 64mb + 64mb swap file ?

It all boils down to how much memory you actually need.

Having more RAM than needed does not improve performance, having less RAM than needed kills performance.

If you need less than 64MB, having more than 64MB will impact performance if the excess is not cached, with no compensating benefits.
If you need more than 64MB, having less than needed will terribly impact performance (modern OS) or give out of memory errors (DOS), and the impact of those will be far worse than the reduction of impact due to L2 uncached RAM.

Key is: there is no 'right' amount of RAM, and just maxing things out does not give you best performance. You need to look at what you are doing and tailor your configuration to what it needs. However I fully agree with the point above that almost no sensible applications for an So7 CPU will require 64MB or more. That's why chipset vendors at the time didn't think the cost of supporting more cached memory was worth it. Consider that at the very end of the So7 era in 1999, a low end system (and by 1999 So7 was very low-end) would still have been shipped with 64MB or less. If you want to run things that need more RAM, you should strongly consider a different platform that is designed for that sort of memory (and with CPUs that can power it).

Thanks everyone!