If I'm correct, the size of the RAG ram "decides" of what amount of ram you'll be able to cache, though many people prefer to use a tag ram as big as the 8 other chips for the cache itself, probably because it was cheaper to buy everything in one row.

The chipset can also "decide" of how big the tag ram will be ... Limiting the amount of cacheable memory, just like TX chipsets are doing.

However increasing the cache memory, even if the memory was already fully cacheable before will still lead to some performance increase. But it won't be as big as getting from 0KB L2 cache to 128/256KB

I haven't come across exactly what you are looking for, but the chart which is available is the maximum amount of RAM per amount of cache so that all the RAM is still cached. Of course you must increase the TAG RAM size depending on the amount of cache.

For direct-mapped cache, generally,
256 MB needs at least 1024 KB if cache set to write-thru mode
128 MB needs at least 1024 KB if cache set to write-back mode
128 MB needs at least 512 KB if cache set to write-thru mode
64 MB needs at least 512 KB if cache set to write-back mode
64 MB needs at least 256 KB if cache set to write-thru mode
32 MB needs at least 256 KB if cache set to write-back mode
32 MB needs at least 128 KB if cache set to write-thru mode
etc...

In most cases, write-back L2 cache has little benefit, 2-5% at best. CPUMark99 seems the most sensitive to L2 cache type.

In an attempt to answer the original question, I think the consensus is that if, for example, you have 4-16 MB of RAM, there is little-to-no recordable benefit in going beyond 256 KB of L2 cache. Going beyond 256 KB is usually for satisfying the memory cache range.

It would be interesting, though, to see a benchmark study done in a system with, say, 8 MB of RAM, and increasing the cache from 64K, to 128, 256, 512, and up to 1024K. And to note which benchmarks benefit the most from the increasing cache size.

256 is my recommendation. I've got pretty much all boards configures with that in mind. It's a good balance of size and speed.

jesolo, I just remembered that I did a comparison somewhere on vogons which shows that for a fixed amount of cache, say, 1024K, using less RAM did not improve benchmark results. There were a few other things tested for. Can't remember where it is now, but it was mixed into some other thread.

feipoa wrote:

It would be interesting, though, to see a benchmark study done in a system with, say, 8 MB of RAM, and increasing the cache from 64K, to 128, 256, 512, and up to 1024K. And to note which benchmarks benefit the most from the increasing cache size.

This has already been done and it shows performance increases a bit with larger L2 cache no matter the memory size.

Like here:
Re: Phil's Ultimate VGA Benchmark Database Project

Byte Magazine, volume 14, issue 3 (March 1989) [archive.org]

Awesome article on page 281 for everything you ever needed to know about cache RAM.

kixs wrote:

This has already been done and it shows performance increases a bit with larger L2 cache no matter the memory size. […]
Show full quote

feipoa wrote:

It would be interesting, though, to see a benchmark study done in a system with, say, 8 MB of RAM, and increasing the cache from 64K, to 128, 256, 512, and up to 1024K. And to note which benchmarks benefit the most from the increasing cache size.

This has already been done and it shows performance increases a bit with larger L2 cache no matter the memory size.

Like here:
Re: Phil's Ultimate VGA Benchmark Database Project

Yes, that is the link. Could you provide a graph which shows benchmark result as a function of cache size for 64, 128, 256, 512, and 1024K? Ideally, a wide spread of benchmarks is ideal to get an idea for ave, min, and max observable benefit. A good starting point would be at least the following: pcpbench, 3dbench, doom, quake, sysinfo, speedsys, landmark, cpumark99, wintune98, sandra99...

Thanks to everyone who has already replied to this thread - this is definitely helpful to me as well, since I've always wanted to increase the L2 cache on my 386 & 486 motherboards, but wasn't exactly sure what the optimum configuration was.

I like to max out every aspect of every board. If it can take 1024K, it will have 1024K.

feipoa wrote:

kixs wrote:

This has already been done and it shows performance increases a bit with larger L2 cache no matter the memory size. […]
Show full quote

feipoa wrote:

It would be interesting, though, to see a benchmark study done in a system with, say, 8 MB of RAM, and increasing the cache from 64K, to 128, 256, 512, and up to 1024K. And to note which benchmarks benefit the most from the increasing cache size.

This has already been done and it shows performance increases a bit with larger L2 cache no matter the memory size.

Like here:
Re: Phil's Ultimate VGA Benchmark Database Project

Yes, that is the link. Could you provide a graph which shows benchmark result as a function of cache size for 64, 128, 256, 512, and 1024K? Ideally, a wide spread of benchmarks is ideal to get an idea for ave, min, and max observable benefit. A good starting point would be at least the following: pcpbench, 3dbench, doom, quake, sysinfo, speedsys, landmark, cpumark99, wintune98, sandra99...

pcpbench, 3dbench, doom, quake are already benchmarked with 256/512/1024kb cache sizes in the linked post. Never got around to do other benchmarks - especially Windows.

As others have said, 256KB is the sweet spot. Anything higher and the improvements are measurable, but insignificant. Probably ~5% or less in most cases.

feipoa wrote:

Of course you must increase the TAG RAM size depending on the amount of cache.

In terms of the above, most motherboards I've come across, seems to specify a tag cache RAM size of 8k x 8 for 128KB of cache and then increase the tag cache RAM size to 32k x 8 for 256KB and above.
Does this sound correct?

jesolo wrote:

feipoa wrote:

Of course you must increase the TAG RAM size depending on the amount of cache.

In terms of the above, most motherboards I've come across, seems to specify a tag cache RAM size of 8k x 8 for 128KB of cache and then increase the tag cache RAM size to 32k x 8 for 256KB and above.
Does this sound correct?

The size of tag ram depends on the implementation of L2 cache. Sounds like your board implements direct mapping scheme with 16 byte cache lines.

kixs wrote:

pcpbench, 3dbench, doom, quake are already benchmarked with 256/512/1024kb cache sizes in the linked post. Never got around to do other benchmarks - especially Windows.

It would be interesting to see the transition to asymptotic performance gain. With 256-1024K only, you are only showing the asymptote. Could you take this down to 128 and 64 K to show this effect and include a few more benchmarks on graphs?

feipoa wrote:

I like to max out every aspect of every board. If it can take 1024K, it will have 1024K.

you and me are cut from the same cloth.

All of my systems are topped out (well mostly), except my obsession with the 486 DX4-100 out of nostalgia for that chip because it's been in the most stable 486 systems I've ever built (not to mention the ones that put up with the most of me stressing them out with software designed for Pentium or better hardware).

Stressing out a 486 with Pentium-intended software; there is no greater joy!

feipoa wrote:

Stressing out a 486 with Pentium-intended software; there is no greater joy!

What's nuts about this one I'm working on is she seems ready for multitrack recording now - updated the Hard DIsk controller driver to the Winbond VL-IDE I am using and now n-track and QuartzAudiomaster seem to be able to keep up.....nutz - a 486 powered DAW with a 20GB HDD....but the real truth comes when I run a guitar into it.