If it runs fast enough for your needs without cache, then you don't need cache.

If you need it to run faster, then you should swap the boards out.

What is the question here?

You ain't missing much with L3 cache, at least as far as games are concerned. It's troubling that it's failing the tests though, have you increased the FSB over 100MHz? Why else would the cache not be coping?

no its running 4x100mhz, my K6-3 is overclocked, but I set it back to its defaut of 3.5x95 as well and the cache fails, I even set the board to 66mhz FSB and the cache still fails. 🤣

Did the cache fail if you tested with the K6-2 ?
It may be that it has a problem operating as level 3 - if there is a BIOS update, it may help
ftp://ftp.fic.com.tw/motherboard/bios/socket7/pa-2013/ - not sure about versions though

With the K6-3 disabeling the L3 cache only lowers the performance with ~1%, there are even cases where you gain performance without L3 with everything else the same.

Skyscraper wrote:

With the K6-3 disabeling the L3 cache only lowers the performance with ~1%, there are even cases where you gain performance without L3 with everything else the same.

Pretty much this. I have the same board and the L3 makes very little difference, but that's only if L1 and L2 caches are enabled. If one of those gets disabled, then L3 will make a much more noticeable impact.

And the latest BIOS version is JI4333, which does support the + series of the K6s.

its not a + cpu, and I never tested it with the K6-2, I got this board and chip together from a trashed computer in a dumpster, so the K6-2 400 has never touched it.

well double checked while i was at work today, I ran tests on cache last night, before I left i disabled the cache, and its been running memtest for the past 10 hours, no errors, so definatly the cache.

Skyscraper wrote:

With the K6-3 disabeling the L3 cache only lowers the performance with ~1%, there are even cases where you gain performance without L3 with everything else the same.

guess you were using an ali5 board
but via boards live on cache, they simply fall to eat dust if onboard cache is disabled. you can expect a 10~15% performance drop even if you are using k6-2+/3+, and over 30% for k6-2 without integrated cache.

You are correct, the 1% figure was with Gigabyte GA5-AX

I have not seen 30% performance degradation with (PC Chips) VIA MVP3 and K6-2 when running more memory than the boards can cache but the performance hit is pretty nasty. I have never tried to disable the L3 cache with a K6-3 using my PC Chips M577 boards as they do not support FSB over 100 MHz and the cache seems to work great at that speed.

Skyscraper wrote:

You are correct, the 1% figure was with Gigabyte GA5-AX

I have not seen 30% performance degradation with (PC Chips) VIA MVP3 and K6-2 when running more memory than the boards can cache but the performance hit is pretty nasty. I have never tried to disable the L3 cache with a K6-3 using my PC Chips M577 boards as they do not support FSB over 100 MHz and the cache seems to work great at that speed.

i have benchmarked that. running k6-2-550 and mvp3 board with 2mb cache, superpi timed at 6:50, and 10:15 with onboard cache disabled. thats exactly taking 50% more time, phew.

A range of benchmarks might be of interest. To see what types of programs are mostly affected. Might check this out with a VIA board and see how it affects popular DOS and Windows gaming benchmarks.

The K6-3 has 256kb of on-board L2 cache running at CPU frequency - as much as a socket 370 pentium 3 or an Athlon XP. You don't need the motherboard's cache witch runs at FSB frequency - it might actually slow your system down. I'd disable external cache completely.

kanecvr wrote:

The K6-3 has 256kb of on-board L2 cache running at CPU frequency - as much as a socket 370 pentium 3 or an Athlon XP. You don't need the motherboard's cache witch runs at FSB frequency - it might actually slow your system down. I'd disable external cache completely.

In one of my video reviews, I actually looked into this. I found it not to be true / a myth, that disabling motherboard cache would increase performance.

This was on a GA-5AX though.

by how cache is structured, the CPU will first attempt to fill the L1, when it gets full it will fill the L2, anything that wont fit into either of those goes into the L3, anything that wont fit into the L3 will get stored on ram. The L3 Cache is quicker than paging directly to ram, its a faster access. Disabling the Motherboard Cache should make a noticable impact on preformance. Its part of why a K6-3 in office applications was faster than a Pentium 3, and why a K6-3 500 is almost as fast as An Athlon Classic in pure interger work, it has a larger cache, but the interger unit is very similar to the K6, I would prefer to keep the cache on, but I can't get it stable.

@candle_86: Not quite. L1, L2 and l3 are filled at the same time.Depending on cache affinity and LRU strategy it is possible that L2 and L3 don't always contain the same data even if they are the same size. And Athlon was the first x86 CPU to use a victim cache, ie. data was at first only cached in L1, not in L2, and when it was flushed from L1 it was then stored in L2. Thus the cache capacities add up instead of L1 being included in L2.

Yes I'm aware on a K6 class just like an Intel from the same time total cache area is only 1024k, all data in L1 is mirrored to L2 and L3. But the CPU still puts the most frequent stuff in L1 then mirrors it in the other 2 caches. But L2 can hold data not in L1, and L3 can hold data not in the other two options.

Free Space per cache

L1 = 32kb Data, 32kb Instruction

L2 = 192k after mirroring L1

L3 = 768k after mirroring L2

But its still faster than going to Ram

Thought I'd expand on this a little bit. I did some benchmarks comparing performance using Phil's benchmark stuff with L3 enabled and disabled (L1 and L2 always enabled). Using a K6-III+ 400MHz, 100MHz FSB, Matrox Mystique 220, 512MB CL2 PC100 RAM, and the FIC PA-2013 motherboard which uses VIA's MVP3 chipset.

This is arranged with L3 on (left) vs L3 off (right).
3DBench: 286.2 fps vs 286.2 fps
PCBench: 122.1 fps vs 110.1 fps
Doom: 710 realtics (~105.20 fps) vs 722 realtics (~103.45 fps)
Quake: 61.5 fps vs 57.5 fps

So there is a little dip in performance, depending on the application, but it's really not too much, with 3DBench not even changing between the two.

Great work in checking this out!