VOGONS


The Ultimate 686 Benchmark Comparison

Topic actions

Reply 20 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++
remax wrote:

Do you think linear interpolation would be valid to get the results of a K6-3 450?

(either with other K6-3 or on the average gap of performance beetween K6-2 or K6-3+?)

You mean for results on a non-super7 (75x6.0)? I already ran the K6-III-450 on a super7 at 4.5x100.

Plan your life wisely, you'll be dead before you know it.

Reply 21 of 145, by remax

User metadata
Rank Newbie
Rank
Newbie

As far as i know, K6-III and K6-3+ are not exactly the same processor, are they?

As there is a gap of performance between K6-II and K6-2+, i expect the same between k6-III and K6-3+

I am not expert, so i might be wrong, hence why i am asking...

Reply 22 of 145, by Tetrium

User metadata
Rank l33t++
Rank
l33t++
remax wrote:

As far as i know, K6-III and K6-3+ are not exactly the same processor, are they?

As there is a gap of performance between K6-II and K6-2+, i expect the same between k6-III and K6-3+

I am not expert, so i might be wrong, hence why i am asking...

The K6-2 doesn't have any L2 cache while the K6-2+ has 128KB L2 cache.
Both K6-3 and K6-3+ have 256KB L2 cache, so should perform identical (at the same clockspeed).

Edit:K6-3+ is basically a die shrunk K6-3 and consumes less power.

Whats missing in your collections?
My retro rigs (old topic)
Interesting Vogons threads (links to Vogonswiki)
Report spammers here!

Reply 23 of 145, by remax

User metadata
Rank Newbie
Rank
Newbie
Tetrium wrote:

The K6-2 doesn't have any L2 cache while the K6-2+ has 128KB L2 cache.
Both K6-3 and K6-3+ have 256KB L2 cache, so should perform identical (at the same clockspeed).

Edit:K6-3+ is basically a die shrunk K6-3 and consumes less power.

Thanks 😀

Reply 24 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++
feipoa wrote:
remax wrote:

Do you think linear interpolation would be valid to get the results of a K6-3 450?

(either with other K6-3 or on the average gap of performance beetween K6-2 or K6-3+?)

You mean for results on a non-super7 (75x6.0)? I already ran the K6-III-450 on a super7 at 4.5x100.

Too bad. I was actually hoping someone was looking for results from a K6-III-450 on an i430tx board at 75 x 6.0. I only ran the K6-III on the i430tx at 133, 300, 333, and 500 MHz. Ironically, after spending a year casually benching all these 686 CPUs, I decided to setup a K6-III-450 on an i430tx (6.0 x 75), for which I do not have tabulated results for.

In general, the i430tx seemed to perform slightly better than Super7 board on an FSB-to-FSB basis. A Super7 board at 4.5 x 100 will still beat the i430tx though. For an AMD K6-III-500 at 83 x 6.0, the performance was 3.5% better on the i430tx compared to the MVP3 Super7. However, the Super7 at 100 x 5.0 was 1.3% better than the i430tx (83 x 6.0). The Pentium 262 MMX was also 2.3% better on the i430tx. For some CPUs, the Super7 was faster clock-for-clock/fsb-to-fsb. These CPUs were the Cyrix MII and original K6.

remax wrote:

As far as i know, K6-III and K6-3+ are not exactly the same processor, are they?

As Tetrium pointed out, the K6-III and K6-3+ are the same performance-wise (clock-for-clock). The K6-3+, however, seems to overclock better than the K6-III. Also, some Super7 boards may only work with the K6-III instead of the K6-3+. Certain revisions of the Asus P5A-B is one such example.

Plan your life wisely, you'll be dead before you know it.

Reply 26 of 145, by sliderider

User metadata
Rank l33t++
Rank
l33t++

Has anyone successfully overclocked a Winchip C6? I always thought they were like the Cyrix 6x86, already running pretty close to their thermal limits out of the box. Can a 240 (60 x 4) go to 266 (66 x 4) or is that too much?

Reply 27 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++

I no longer recall exactly how far I tried to take the original C6, but I think it didn't go much past its rated frequency. The 233 MHz rated Winchip2 seemed to run OK at 262 MHz.

Plan your life wisely, you'll be dead before you know it.

Reply 28 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++

New PDF report and select PNG images were replaced.

This change affects IBM 5x86 100/120/133, AMD X5-133/160, and Intel DX4-133 values for Quake 1, FPU, and Overall scores. When first tabulating these marks, I inadvertantly had DOS sound enabled. The changes indicate the corrected scores with sound disabled.

EDIT: I reverted the files back after realising they were right the first time. The increase in Quake 1 scores noted recently was because I fogot the Rage 128 was still in the computer instead of the desired Matrox G200.

Plan your life wisely, you'll be dead before you know it.

Reply 29 of 145, by subhuman@xgtx

User metadata
Rank Oldbie
Rank
Oldbie

Props for making this thread possible.

By the way, Could you add in a future the same P1 mmx but running at 292 (83x3.5)? I'd like to see how it compares to k6 chips at that speed.

7fbns0.png

tbh9k2-6.png

Reply 30 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++

I beleive I tried this on the 430TX board, but I couldn't get it stable past 262 MHz. You'll have to use the 300 MHz results for the P1 MMX on the VIA MVP3 board as an approximate substitute. It is unlikely I will make any additions to this comparison anytime soon, but when/if I do, I'll try to add the 3DNow! patch for Quake on Winchip and AMD CPUs.

EDIT: I may also add an AMD X5-133 at 2x66 MHz. Actually, I've already taken all the data for this configuration.

Last edited by feipoa on 2013-07-08, 12:48. Edited 1 time in total.

Plan your life wisely, you'll be dead before you know it.

Reply 32 of 145, by gerwin

User metadata
Rank l33t
Rank
l33t
feipoa wrote:

* FastVid was not enabled for testing. Enabling FastVid may increase highly graphic DOS-based benchmark scores, such as 3Dbench and Doom, when using a Pentium Pro, Pentium II, or Pentium III CPU.

How about a 270% increase in frame rates in 640x480x8. It makes a huge difference. Seems pointless not to use this added power. AMD, Cyrix and Centaur also supported Write combining at some point.
FastVid does the job but I found it inferior compared to MTRRLFBE. Neither tool remains resident. The latter can easily be ordered to leave 320x200 VGA mode unaccelerated, for compatibility sake. VGA mode is peanuts for this generation of CPUs anyways.

The earlier benchmark run by swaaye, 133 MHz Challenge, had write combining enabled. Unfortunately the images are no longer visible. I did write down the results back then, maybe I should put them in an excel sheet and upload them here. Edit: Just did that and attached it on page 4 of that topic.

Reply 33 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++

I had a topic going which asked about what people would like to see in the 686 benchmark comparison, but I received little feedback. To go back and patch these results for all CPUs tested would be very time consuming. Maybe if I ever re-test the 3DNow! chips for Quake, I'll add a row for w/FasVid results and w/out. With thousands of bricks in the wall, there may be a few bad ones - don't forget about all the good ones!

Plan your life wisely, you'll be dead before you know it.

Reply 34 of 145, by gerwin

User metadata
Rank l33t
Rank
l33t

On second thought, it is not that I have a problem with the benchmarks. They are very interesting and much appreciated. At the moment I was thinking it gives undue credit to some of the older CPU's that do not support write combing. And that the FastVid remark of the article was rather limited.
I hope my post will help to put it in perspective, when selecting one of these CPU's for >320x200 gaming, even though it sounded a bit harsh I guess.

--> ISA Soundcard Overview // Doom MBF 2.04 // SetMul

Reply 35 of 145, by kool kitty89

User metadata
Rank Member
Rank
Member
luckybob wrote:

man, Pentium pro overdrives are AWESOME chips... Something doesn't sit right with me on the quake 1 results though. I'm going to run that once more and double check. for some reason not only are those chips WINNING but beating chips 2x as fast in mhz. honestly i'd expect to see them just under the p2 xeons.

unless quake 1 dos loves p2-overdrives...

I'm really wondering about this too. Those PII Overdrive figures seem REALLY high relative to PII/PIII/Xeon. (they're essentially the same core architecture, just with different cache and FSB arrangements, as such I'd have thought the PII Overdrive would be a little below PII Xeon of similar clockings)

Though it's also interesting that the MII fared almost identially to the K6 and K6-2 at similar FSB/core clock speeds. (with the weaker FPU, I'd have thought the K6 would have more of an edge . . . then again, Quake 1 is REALLY P5 optimized, and the M2 might end up catering more to the programming quirks of P5 than the K6 does)

On another note, I just noticed how amazingly fast the Cyrix CPUs are at Doom, though the K5 does even better anbd K6 fairs pretty well. The P5 based chips make out pretty well too, behind the AMD and Cyrix offerings, but still in the ballpark. The P6 core based chips do really bad there though, at least compared to the typical integer performance for later benchmarks. (the PII overdrives are way down there too, in stark contrast to the Quake 1 trials)
The Winchip doesn't do great either, but it's not terrible considering it's closer to a Cyrix 5x86 or 486 internally (or Media GX -though the GX itself is really I/O bound), but the VIA Cyrix/C3 chips seem to be the worst by far.

It's really interesting to compare the pre-pentium optimized software in general, among other things it points out more how the later chips did their own execution optimization internally, rather than relying more on compiler/programmer optimized scheduling. (though it also points out how well performnace was improved in the specific areas commonly used in old software -granted, integer computationally intensive 3D/pseudo 3D games compared to old business apps and such, or plain 2D games)
Too bad games like X-Wing, TIe Fighter, or Wing Commander III aren't convenient to use for benchmarks like Quake or Doom. It would be interesting to see how things compared for the last generation of non-pentium non-FPU specific top-teir polygonal 3D games.

On the note of the C3 being weak . . . in general it seems really poor for both multimedia/game and general performance. It really makes me wonder why VIA even chose that chip over the MII based projects (like the Cayenne core used in the original Joshua Cyrix III) brought over from Cyrix. WIth the Joshua design apparently having somewhat better clock scaling than older MII derivatives combined with the big L2 cache, much improved FPU, and added 3DNow! it should have easily managed better than 2x (maybe more than 3x) the per-clock performance of the C3 based stuff. Hell, and if yields, power consumption, and die sizes were a problem, cutting back the cache size should have addressed that a good bit. (given how well the old MII already fares next to the C3, and given the C3's modest cache itself, cutting down the Joshua core's cache should have been pretty reasonable)

There was also the WIP Jalepeno (M3) core design Cyrix was working on before the VIA buy-out, which might have made a really nice follow-on had VIA not cut that team in their downsizing. (granted, a risk given the incomplete nature of the project, but had it eventually made it to market, it should have been massively more capable than the Centaur based designs and more on the level of K7, and like the M2 in terms of IPC rate but with other improvements and MUCH more emphasis on clock scalability, including an 11 stage pipeline -which is one reason it sometimes gets confused with the WinchipIV based design VIA eventually did use)

Reading old articles from '98 and '99 about that NS/Cyrix stuff, aside from the typical "NS bureacratic management bloated and ruined Cyrix" there's the more specific issue of how heavily they seemd to be investing in SoC designs. None of that seemed to pay off, and with the lukewarm success of the Media GX itself, it really makes me wonder why they retained such a high emphasis on that end of R&D. (as it is, it seems kind of a waste that the updated 5x86 core in the Media GX didn't get worked into a S7/SS7 part as a lower-cost complement to the M2 line, and benefitting from the full bandwidth and L2 cache of S7/SS7 -better balanced ALU+FPU+MMX performance than M2 for much better multimedia/gaming performance relative to manufacturing cost, power consumption, or PR rating -which would actually be LOWER than the clock rate like the old 5x86 😉 . . . plus yields and clock speed scaling probably would have fared better than the full M2)

Reply 36 of 145, by feipoa

User metadata
Rank l33t++
Rank
l33t++
gerwin wrote:

On second thought, it is not that I have a problem with the benchmarks. They are very interesting and much appreciated. At the moment I was thinking it gives undue credit to some of the older CPU's that do not support write combing. And that the FastVid remark of the article was rather limited.
I hope my post will help to put it in perspective, when selecting one of these CPU's for >320x200 gaming, even though it sounded a bit harsh I guess.

gerwin,
1) Please let me know exactly which of the employed tests and CPU's are affected by not having fastvid enabled.
2) If you are willing to rewrite my 'limited' remark, please let me know exactly what you want said and I'll copy/paste it in.

It took 1 year to tabulate all this data. By the time I was finished, I didn't have any mood for writing up a proper review. I did note that the comments posted to this thread will serve in lue of a formal review.

Plan your life wisely, you'll be dead before you know it.

Reply 37 of 145, by gerwin

User metadata
Rank l33t
Rank
l33t

Thanks for your suggestion. I will need to test the MTRR/Write combining support whenever possible though. Contrary to just taking the list from the wikipedia article and assuming it works the same as a Pentium II/III.

--> ISA Soundcard Overview // Doom MBF 2.04 // SetMul

Reply 38 of 145, by kool kitty89

User metadata
Rank Member
Rank
Member

A bit off topic, but more in line with the Cyrix comments above about relative performance and such, and the National Semiconductor mess. I hadn't really thought about it before, but aside from the whole management mess and shift in R&D priorities after the Cyrix/NS merger, they also lost the IBM manufacturing connection, and missed out on IBM's fast 20 nm process chips with copper interconnect (and then the 180 nm copper in late 1999) and potentially a good bump in clock speed over the 250 nm IBM/NS parts, or the 180 nm NS parts for that matter. (and that much more potential for the late-gen M2 core, including the upgraded Cayenne/Jedi derivative, and bought more time for the M3 project -aside from the potential of more focused R&D on the desktop CPU platform rather than the excessive investment they ended up puting into SoCs)

Reply 39 of 145, by kool kitty89

User metadata
Rank Member
Rank
Member

It's kind of interesting to compare these old CPU scaling game performance figures on Tom's Hardware:
http://www.tomshardware.com/reviews/3d-accele … step,51-17.html

The 6x86MX and (especially) K6-233 seem to fare better against the P55C and PII in those games (Turok, GL Quake, and Quake II) than most of feipoa's tests. Including the K6-233 actually performing significantly better per-clock for some cases than the MMX-200. (and at least somewhat faster in all cases)

That, and GL Quake II seems to do a fair bit better on those Voodoo Is than the G200. :p