melbar wrote:

Conclusion:

As i did not expect: The K6-2 @171MHz is even more slowly during the disable of L1 and L1+L2 caches (for the 486 and 386 mode), compared to the Pentium @137MHz or @103MHz. That means the Pentium is really faster regarding the speed per clock cycle compared to the K6-2. This has also something to do with the different build of these CPU's: the Pentium has a CISC core and the AMD has a RISC core with x86 to RISC86 decoder. The AMD structure is more influenced by disable the caches.

For the Pentium III @431MHz (66MHz FSB): I am really surprised about the behavior of the coppermine when disable the cache and comparing the two benchmarks. First, it doesn't matters if you disable only L1 cache or both L1+L2 cache. The results are the same. With the 3DBenchmark i've got nearly twice the value of the Pentium@137MHz with L1 disable. But the PCPBenchmark, the Pentium III undermatches all the settings of Pentium & K6-2. Really interesting, i don't know why this can happen...

I think the Coppermine becomes slower with the caches disabled compared to the s7 stuff (especially Pentium 1) as Intel probably made changes to that CPU so it would clock higher (that was a big selling point in those days) and they might've optimized it more to make it also perform good (they kinda did this the right way where VIA accepted too many compromises with it's C3), but I'd reckon that as Coppermine is more optimized, having 1 thing out of order (like its caches) will make it break its performance more.

Perhaps it's something along those lines. Coppermine needed to clock higher yet still perform, but apparently it needs its internal cache to still achieve performance. That would be my first guess.

I'll read your other findings in a minute, well done 😀

edit: And I find the vids Phil (and others!) are making very interesting. Not all of them, but many! 😜

P2/P3 also disable L2 cache when disabling L1 cache. This is also true for K6-3. For S7/SS7 CPUs without L2 cache the L2 cache is on the mainboard and is not automatically disabled with the L1 cache.
IIIRC from results previously posted K6 CPUs do slow down more than Pentium CPUs when disabling the L1 cache while Cyrix CPUs slow down less than Pentium CPUs and P2 CPUs slow down the most (286 level).
K6-2+/3+ are indeed useful because the multiplier can be changed at runtime and down to 2x.

I wouldn't hold any of these points against the CPUs. After all the normal mode of operation is with enabled caches and that's what counts. Disabling caches is for troubleshooting/debugging, and while different CPUs slow down more or less in that case it doesn't make them bad or slow per se, just more or less suited for playing old games of a certain period.

Nice!

A lot of data to digest, I'll make myself some hot chocolate 🤣

EDIT:

Well that is some really nice work you did there and match my own findings. Apart from the Pentium III, but more on that later.

Like what you found, CPU architecture of Socket 7 processors, determines "how slow" the CPU gets with caches disabled. While I don't have (m)any exotic processors, this is what I found:

K6-2: Slowest speed with caches disabled
Pentium: Faster than K6-2 with caches enables
Pentium MMX and Cyrix: Fastest with caches disabled. They will give you a solid 386DX-40 equivalent experience with Wing Commander running too fast.

You can also tweak the speed a tiny bit by changing memory type (I have these EDO sticks in SDRAM format, they are slower than normal SDRAM) and you can also change the timings in the BIOS.

Now back to the Pentium III and any Slot 1 processor that I tried.

Firstly, it's easiest to think of L1 BIOS settings as CPU cache. And L2 BIOS settings as motherboard cache.

Slot 1 boards don't have motherboard cache, so the L2 setting doesn't do anything. I'm not sure why the BIOS on such a machine have a L2 option to be honest.

Anyway, what I found is that Pentium III is either super fast, or super slow when cache disabled. With super slow I mean 286 speed, so slow that Indiana Jones and the Last Crusade runs slow 😒

So I find that the cache tricks works great on 486 and Pentium machines, but Pentium II and higher I don't use it and just recommend these for newer DOS games.

Tetrium wrote:

I think the Coppermine becomes slower with the caches disabled compared to the s7 stuff (especially Pentium 1) as Intel probably made changes to that CPU so it would clock higher (that was a big selling point in those days) and they might've optimized it more to make it also perform good... ..., but I'd reckon that as Coppermine is more optimized, having 1 thing out of order (like its caches) will make it break its performance more.
Coppermine needed to clock higher yet still perform, but apparently it needs its internal cache to still achieve performance.

Yes, i agree to that. The PIII was the next step for Intel around the turn of the century. When the data on wiki is correct, they have following different stage pipelines:
Pentium (P5 gen.) - 5 stage pipelines
Pentium Pro (P6 gen.) - 14 stage pipelines (and i think PII has the same basis)
Pentium III (P6 gen.) - 10 stage pipelines
It's both, the combination of longer pipeline, and the more tiny fabrication process that they (PII + PIII) performs higher, but regarding the point of deactivate the caches, the PII & PIII 's it seems even worse for their execution unit.

idspispopd wrote:

P2/P3 also disable L2 cache when disabling L1 cache.
I wouldn't hold any of these points against the CPUs. After all the normal mode of operation is with enabled caches and that's what counts. Disabling caches is for troubleshooting/debugging, and while different CPUs slow down more or less in that case it doesn't make them bad or slow per se, just more or less suited for playing old games of a certain period.

Ok, that means you have no control for the PII and PIII with the bios setting "external cache" which imply to be the L2 cache.
Well, sure it's no normal mode when you disable the caches, but for me it's something like the turbo button of the 386 and 486's. I think it's a good possibility to run 'ancient' games not in software mode but with real hardware like 'tractor-pulling' 😁
Apart from the additional money you have to spend for old hardware, it's good to have the control of speed in that way. With the K6-2, i have now a rage of ~386DX-20 to full 500Mhz speed (PII equivalent). With the PIII, i've got a range of 433-866Mhz, and also whith the third old companion i've used in the olden days (AthlonXP) has a range of 1,33Ghz(1500+) to max. 2,16Ghz(2700+).

PhilsComputerLab wrote:

Nice! […]
Show full quote

Nice!

A lot of data to digest, I'll make myself some hot chocolate 🤣
EDIT:
Pentium: Faster than K6-2 with caches enables

They will give you a solid 386DX-40 equivalent experience with Wing Commander running too fast.

Thanks 😎
Did you mean " K6-2 with caches disabled" instead of enables? Cause when i compare my values of the K6-2@ 171MHz with the 166's Pentiums' and MMX's from your ultimate VGA benchmark chart, then i have nearly the same.
The problem with WC, i totally agree. I was afraid when i saw your video with the DX-40. During my test's i've tried all the scenarios with different clock+ disables caches to see where's the border. It's not easy, but when i was a child, i've played thru this amazing game at the low speed of 7.16Mhz of a MC68000 processor 😊

PhilsComputerLab wrote:

Anyway, what I found is that Pentium III is either super fast, or super slow when cache disabled.

Well, i have anyway not a clear idea, that for example the 3DBench runs with 486DX66 speed whereas with the same settings the PCPBench is a really low 386 / or fast 286...

Hello! I seem to remember that there used to be a thread a while ago where people had made a comparison table between different 386 and 486 speeds and K6-2/3 CPUs with cache disabled. It was basically an easy way to find out the settings you had to use on your K6 CPU to achieve a specific speed. If I wanted to do 386DX40 for example, I had to disable both caches and use a specific speed.

Is there anything like this available somewhere? I want to "emulate" a 486/DX2 66 using a K6-III+, but I want to be sure I've hit that sort of performance.

This thread?

Let's benchmark our systems with caches disabled

Well done, melbar. That's a lot of work! And thanks for sharing with us. 😀

melbar wrote:

Well, i have anyway not a clear idea, that for example the 3DBench runs with 486DX66 speed whereas with the same settings the PCPBench is a really low 386 / or fast 286...

I think what you're seeing here is 3dbench 1.0c not scaling down. In that case, try 3dbench 1.0 (the older version) which will give more accurate results when the cpu gets into the slow 386/fast 286 range.

melbar wrote:

This thread?

Let's benchmark our systems with caches disabled

No, that wasn't it, but it did help me out, so thank you very much! 😀