K6-III + 400MHz (1MB L3 cache motherboard)
Quake 2 testing:
Cache L1 Cache L1 Cache L3 FPS
ON________On______On_______60.1
ON________On______Off_______58.2
ON________OFF_____OFF______52.9
OFF_______OFF_____OFF______0

Pentium MMX 233 (1MB L2 Cache motherboard)

My pong game:
Cache L1 Cache L1 FPS
ON_______ON_______1813
ON_______OFF______1775
OFF______ON_______305
OFF______OFF______121

Pentium MMX 250 (1MB L2 Cache motherboard, Overclocked)

My pong game:
Cache L1 Cache L1 FPS
ON_______ON_______2030
ON_______OFF______1956
OFF______ON_______434
OFF______OFF______86 (wtf, lost compared to 233)

Doom Benchmark (Pentium MMX 233 and K6-III+ 400)
http://i.imgur.com/Yvk3Mh6.png

Tertz wrote:

link I used it in my bench theme in the signature. May to be useful any benches which were made with 286 in mind, as they run n […]
Show full quote

clueless1 wrote:

Can you put a link to Speedtst? I'm not sure I'm familar with it.

link
I used it in my bench theme in the signature.
May to be useful any benches which were made with 286 in mind, as they run not too long on 286-386 cpus level of wich is expected by switching off the cache1. If a bench runs >15 min - it's bad, and becomes worse if there are several such.

Thanks, I've added speedtst to my results. 😀 Going through the PII at various multipliers and Celeron 333 now. When I publish them you are welcome to pull them out of my results to add to yours.

Almost done with all the Pentium multipliers. Then on to K6-2. One really interesting anomaly is that when L1 is disabled on the Pentium, it is consistently faster at 120Mhz than it is at 133Mhz--like every single benchmark is faster at 120 vs 133 when L1 is disabled.

Should Cache L1 Cache L1 Cache L3 read Cache L1 Cache L2 Cache L3 ?

So here are the results with only the K6-2 remaining.

Surprises so far:
-Sysinfo is surprisingly accurate and consistent when compared with Doom results. The only outlier is the wildly underclocked "Pentium II 133" which Norton thinks is much slower than it really is. Speedsys also gets confused by the Pentium II 133.
edit: Speedsys does NOT get confused, Speedtst does.
-with L1 disabled, the Pentium is fastest at the 60Mhz bus speeds (120Mhz and 90Mhz) but still seems to be equivalent to a slow 386.
-the Celeron and Pentium II are pathetic with L1 disabled--286 speeds.
-with L1 enabled, the Pentium II 333 is slightly faster than the Celeron 333; with L1 disabled, the roles are reversed.

Things will get more interesting with the K6-2. At some point I should create a new thread and ask for volunteers to submit results. I'd be looking for stock speeds with and without L1 cache. No overclocked buses, etc. If you have an unlocked multiplier, feel free to submit results for multiple "stock" speeds (like my Pentium and Pentium II results, each done with one cpu with unlocked multiper).

I'm considering removing the Fastvid columns, I'm not sure they help much with what I'm trying to show. Thoughts?

clueless1 wrote:

Going through the PII at various multipliers and Celeron 333 now. When I publish them you are welcome to pull them out of my results to add to yours.

Besides speedtst in real DOS, my table contatins 4 numbers gotten on DOSBox. In case you have fun for more testing, would be interesting to compare Celeron 300 (wich exists in my table) with P2 300 in Win98 by the method described in the bench theme. Data for Celeron was gotten in XP as in Win98 results were significatly lower in DOSBox, do not know why. Is the situation same on normal P2 in Win98 is interesting to check.

Tertz wrote:

clueless1 wrote:

Going through the PII at various multipliers and Celeron 333 now. When I publish them you are welcome to pull them out of my results to add to yours.

Besides speedtst in real DOS, my table contatins 4 numbers gotten on DOSBox. In case you have fun for more testing, would be interesting to compare Celeron 300 (wich exists in my table) with P2 300 in Win98 by the method described in the bench theme. Data for Celeron was gotten in XP as in Win98 results were significatly lower in DOSBox, do not know why. Is the situation same on normal P2 in Win98 is interesting to check.

You got it Tertz. I'm just not sure I'll ever have Win98 on the P2 or Celeron. My current Win98 box is with a P3-933 and I don't foresee making another Win98 install.

BTW, I did notice in my benchmarks that Speedtst was very similar to Norton Sysinfo, right down to getting confused about the P2-133. The only synthetic benchmark that wasn't confused by the underclocked P2 was Speedsys.

I've been thinking about what my goal is--accurately documenting CPU performance with L1 cache disabled, and depending on the CPU, L2 cache as well. Some things I'm considering changing:
-Remove all results with L1 cache ENabled. There's enough benchmark lists with numbers like these already.
-Remove results with Fastvid. When L1 is disabled, Fastvid's advantage goes away (except in 3dbench).
-Focus primarily on Speedsys and Doom. If you notice my underclocked Pentium II score at 133Mhz, it confuses many of the synthetic cpu-based benchmarks. Speedsys is not confused though. Its results are very consistent with Doom all the way down the list.
-Remove Quake. No one who's disabling their L1 cache will probably be thinking about running Quake.
-Remove Sysinfo. Its results are very consistent with Speedtst, so why not use Speedtst, which is open source?

All of these would drastically simplify the table of results and make it easier for people to want to contribute.

I would appreciate opinions. Thanks!

clueless1 wrote:

I'm just not sure I'll ever have Win98 on the P2 or Celeron. My current Win98 box is with a P3-933 and I don't foresee making another Win98 install.

In case your MB is slot1 then it supports P2 cpus.

Got sidetracked in another thread. I don't want to cross-post, so here is the tangent that should've been in this thread:
Which CPU: DOS vs. Windows 98 machine?

Tertz--I'm embarrassed to admit it, but I don't remember if my Dell Dimension 4100 has a slot 1 or socket 370. I'll have to check.

clueless1 wrote:

gdjacobs wrote:

clueless1 wrote:

Something like this. The red values are median results taken from Phil's chart.

test.jpg

Yeah, that should work nicely. As it's data for a real machine and not a composite mean value, you can also include specifics in a footnote. That way people can refer back to the appropriate entry on Phil's table.

Actually, the red results are composite values (medians, though, not averages). edit: what I mean is I used a median() formula, so each result is not necessarily from the same machine, and some results might be a median between two results depending on whether there was an even or odd number of candidates. But I think I like where you're coming from. I could take a real result, that's closest to the median for that CPU, to represent that CPU in this table. I wonder which would be preferred (a real result or a theoretical median).

Well, my background is in physics. We always use mean values and standard deviation because the error sources are almost always statistically random. In your case, the mean should be quite close to the median, unless your data set has both significant deviation and notably non-Boolean distribution.

I'll have to remember that the next time I'm feeling clueless. 😉

I did test a few data points with means and they were close, but there were just a few weird results in the database that I thought medians would be better at filtering those out.

Physicist: "We need more data."
Engineer: "I've got my linear relation. Two points is plenty for me."

Tertz wrote:

clueless1 wrote:

I'm just not sure I'll ever have Win98 on the P2 or Celeron. My current Win98 box is with a P3-933 and I don't foresee making another Win98 install.

In case your MB is slot1 then it supports P2 cpus.

Just checked and it's a socket 370. 🙁

I have a rough draft I'd like to get vetted before starting a new thread. Please take a look, guys. All feedback appreciated.

Red entries are the most typical of each processor type taken from Phil's VGA Benchmark, and are the "guideposts" in the list, telling you where your CPU with disabled cache falls into place. There are some where there were only 1 or 2 entries on the list (mostly the slowest processors), but that really can't be helped, and the data does look like it falls into place nicely.

A few notes:
-I decided to remove Speedtst because a) it offered no additional value over Speedsys, and b) it is a separate download, whereas Speedsys is included in Phil's benchmark zip file, so results will be easier/quicker for participants to submit.
-In trying to keep the chart compact, I only included what I felt were the bare minimum specs. Some of the specs of the guidepost results from Phil's chart have missing or vague entries. I did my best to specify them accurately. If you see a mistake, let me know.
-Are there too many guideposts? Specifically, some of the less common entries like the Cyrix chips--should they stay or be removed?
-I set the ceiling at Intel DX4-100. Too high or okay?

I appreciate any and all feedback!

Do you have results for L1 EN / L2 DIS, L1 DIS / L2 EN, and L1 DIS / L2 DIS? Also, I would love to see results for max stock clock and min underclock for each core (Chompers, CXT, Deschutes, Mendocino, etc)? That way, you can plot the full tuning range per core per cache setting.

gdjacobs wrote:

Do you have results for L1 EN / L2 DIS, L1 DIS / L2 EN, and L1 DIS / L2 DIS? Also, I would love to see results for max stock clock and min underclock for each core (Chompers, CXT, Deschutes, Mendocino, etc)? That way, you can plot the full tuning range per core per cache setting.

Hey there, thanks for the comments.

Only the Deschutes and Mendocino support L2 disabling out of my cpus on the chart. I do not have a full range of results, but I have done enough tests to know that disabling L2 by itself does almost nothing to reduce performance. And since we know that just disabling L1 already makes them too slow (unless you are targeting slow 286-class performance), I didn't bother with L1+L2 disabling (those Doom benchmarks would take forever). But using the "Cache Disabled" column is where the participant would be free to test any of those combinations and report such results. So in the case of a K6 III+, there might be a row of results each for "L1", "L2", and "L1+L2".

The thing I didn't want to get into by including full speed results is the FASTVID debate. In my tests, FASTVID makes almost no difference with L1 is disabled. Basically, I saw a small improvement in 3DBENCH2, and no improvement in any other result. But it obviously makes a big difference at full speed for those CPUs that support it. If we did include full speed results, I'd just as soon indicate to NOT use FASTVID, MTRRLFBE, or SETK6D, as it will really make things unwieldy.

But yes, I do have the full speed and minimum underclock results of *my* benchmarks, just not on this chart. I was trying to keep the chart small enough to fit in one screenshot, since I have not looked yet into how to share a Google spreadsheet. That gets me wondering, would it make sense for this to be on a separate tab of Phil's VGA Chart? Phil, if you are reading, what do you think?

edit: I can tell you that the K6-2 166 (66.6x2.5) results are not much slower than K6-2 266. Doom with L1D is 10.28 fps, vs 11.03. That takes it about from an average 486sx-25 to a slow 486sx-25. And a Pentium II 133 with L1 disabled gets a Doom score of 1.18 fps (63435 realticks).

I'm thinking mainly of ALU performance. If a CPU with handicaps on performs similar to a period machine in a Drystone benchmark, it should handle the timing loop of, say, Wing Commander at very close to the same speed. I doubt FASTVID or MTRR settings will have much bearing in such a situation, although you might include data at some point at the top end for all our Quake fans.

As for clock adjustments, I'd just like to know what the range of adjustment is and what performance ranges are not covered by any given CPU. Adjusting the slower board or package level caches on Socket 7 machines and Deschutes systems may be a way of painting in some of those spots that otherwise cannot be reached.

Next revision:

There's a lot of weirdness with 3DBENCH2 and the Pentium II's with L1 disabled. Especially the Pentium II 133. I did double-check that score and verified it over multiple runs. I think 3DBENCH2 is overly graphics subsystem dependent.

Very nice summary! It looks like the results are mainly sorted to PCPbench or DOOM realticks.

clueless1 wrote:

Only the Deschutes and Mendocino support L2 disabling out of my cpus on the chart. I do not have a full range of results, but I have done enough tests to know that disabling L2 by itself does almost nothing to reduce performance. And since we know that just disabling L1 already makes them too slow (unless you are targeting slow 286-class performance), I didn't bother with L1+L2 disabling (those Doom benchmarks would take forever). But using the "Cache Disabled" column is where the participant would be free to test any of those combinations and report such results. So in the case of a K6 III+, there might be a row of results each for "L1", "L2", and "L1+L2".

You forgot your K6-2 system. Even with Sis530 board you can control L1 and L2 separately, or not?

Your're right that disable L2 cache only, will not change the situation dramatically. But there is something you maybe have to add in the chart if you don't want confuse the reader.
Primary, i've compared your values of P100, P133 and K6-2 166 with my values, cause i did already a comparison in the past (For my RetroPC's Projekt: Comparison of Pentium vs. K6-2 vs. PentiumIII @ lowest speed).

P100 and P133:
I can only compare the 3DBench2 & PCPBench result i've created in the past. What i see directly is:
- with full caches, your 3DBench2 are little higher (+4,8% and 6,1%) whereas my PCPbench values are higher.
- with "you write L1D", you'll get values which are less than the values i'll get with "L1+L2 disable".
Main difference is also the hardware in the backround. You've used the intel 430FX with EDO ram and a V3-2000. I have used the same platform like K6-2, the VIA MVP3 with SD-Ram and a ATI RageXL.

For the AMD K6-2, you've tested all the speed with full caches. You have shown the L1 disable results until the 266MHz where you will get the 486 mode. Here are further results (L1+L2 disabled) for CPU's like the K6-2 also relevant cause you can reach 386.
I've compared your values with mine, but since these results are really matching to each other. I know that you cannot compare them to 100% due to different chipset and VGA, but even with my 3% overclocked FSB, the values are reasonable.

Since you have also DOOM values, i've made a re-bench of my K6-2 171Mhz, now with another VGA (Geforce2-MX400 AGP), including DOOM and also a investigation in SETK6.
I have included results with "full caches" and "L1 on, L2 off" and 3 different runs:
- standard setting, write allocation on, write combining off
- write allocation off, write combining off
- write allocation on, write combining on
These setting have only a effect when you have L1 enabled.

At the end, I've got a strange result for DOOM: with write allocation/ write combining Off, i'll get the best results. I don't know why....?!