Nice work! 😁

swaaye wrote:

I wanted to use K6-2 and K6-III+ as well but I forgot that the 2x multiplier is remapped to 6x on them.

Old non CXT-core (stepping 0) versions of the K6-2 still support the x2 multiplier.
K6-2+/3+ CPUs can have their multiplier altered by software tools. I recommend CTU for Windows and K6CLK for DOS.

Ya my K6-2 400 is too new. Oh well.

Thanks for reminding me about the software multi tweaks for K6+. I forgot about that. Maybe I'll give that a go at some point here.

Damn good thread!

Nice to see that something can take advantage of the bas ass 1mb P Pro 😀

Yup PPro 1MB is no slouch. 😀

Well, I tried the software multiplier tweakers with my K6-III+ but they aren't changing its multiplier. I tried k6dos.sys and k6speed. Something about the BIOS of the ASUS P5A must be messing with things I guess. I am running the beta BIOS that supports K6+ chips so I don't know what's up.

This is what i wanted to ask when i read your previous CPU benchmarks thread! I was particularly interested in the K5/K6 vs pentiums.

I think one more CPU is missing... The Deschutes PII @133MHz.

My 350MHz P2... well, how do i say it, it's the longest ever to ship from an ebay seller! The seller continues communication at the rate of ~0.6 email per week 🤣

I'd like to run a Pentium II, but unfortunately I'm fairly certain that they don't have a 2x multiplier. The lowest is 3.5x for 233 MHz @ 66 MHz FSB.

Well, I discovered why I couldn't get the multiplier of the K6-III+ to change. I actually had the K6-2 installed!!!! OMG! A special moment for me.....

Anyway, K6-III+ @ 133 MHz results are now included in the original post. K6-2 CXT and later cores are superior to the original K6 in various ways. The end result for these tests is improvement primarily in the Sandra multimedia tests with MMX and 3DNow. It is much faster with MMX than the original K6 per clock. Quake benefited only slightly, however. The FPU still stinks.

All this is really interesting. KUDOS to swaaye for getting so many different micro-architectures running at the same speed. I think the biggest shock to myself was how well that PPro 1MB held it's ground against everything else. Looks like it led the Quake2 numbers, even if the synthetic benches put it closer to the middle. Must be correlated to the memory speed, which speedsys ranked the highest. What could we do next.... maybe a Tualitan vs. Athlon vs Athlon XP vs P4 Willamette @ 14.ghz?!?!?!

Per-clock performance of these chips has been something I've been curious about forever. So this was a fun benchmark project for me. 😀

Quake was written basically to leverage the Pentium architecture. This is well documented. They literally looked at the Pentium and wrote the game around its quirks and advantages to make their 3D engine as fast as possible for what they thought would be the most popular CPU. Another thing is that P5 and P6 have the best FPUs of this bunch. They are the only chips with pipelined FPUs and Quake loves that. Quake might be the oldest game that is so FPU intensive. It was rather unique at the time for sure.

AMD didn't put a decent FPU together until Athlon. Winchip is architecturally more a 486 than a 5th gen chip (see how it compares with Am5x86). AMD K5 is a really interesting chip IMO with superb performance, and it is actually entirely different from K6 being based on AMD's older 29k RISC CPU. But the clock speed couldn't ramp up because of its design and so it's flawed and that's why it was dumped.

As to what to do next, I dunno. The newer CPUs are better documented in reviews. We all know a P4 will get punished at the same clock as P3 or Athlon. 😀 P3 Coppermine/Tualatin and Athlon TBird/newer are very closely matched. P3 Katmai is a bit slower than the newer P3s.

Is there a way to have SiSoft Sandra use Standard ALU and FPU in every Multimedia Bench?
Comparing ALU vs MMX, FPU vs 3DNow is quite theoretical and strongly bench-code dependent.

On first glance I would prefer always a comparable basic set there with standard ALU and FPU and after that enable additional specific instruction sets.

I did a little experimentation and found that PII does actually have a 2x multiplier. Undocumented, but it is there. So we now have Pentium II "Klamath" 133 MHz results. 😀 It performs in between the PMMX and PPro per-clock.

Running a PII at such a low clock speed is somewhat strange because it takes the L2 cache clock all the way down to 66 MHz (133/2). Normally a PII has a L2 clocked at 116MHz or higher. Notably, according to Speedsys, PII's 66MHz L2 is rather competitive with PPro1MB's even though there's a 2x clock speed difference!

PII "Klamath" has some strange MMX performance. It looks like the Pentium MMX is better here for some reason.

Also, I figured out how to disable SIMD code in Sandra's multimedia test. I don't think I'm going to go back and retest everything I did earlier, but I posted this one for those who care.

swaaye wrote:

PII "Klamath" has some strange MMX performance. It looks like the Pentium MMX is better here for some reason.

Well overall the Int ALU and 3DNow! / MMX and probably also the Sandra FPU values seem to be dependent on the memory or L2 Cache speed. You can see this f.e. for the K6 vs. K6-III+. At Int MMX the K6 scores 285 whereas the K6-III+ goes up to 434. Though the same CPUs show nearly identical Whetstone values (160/161).
It looks like the Dhry and Whet values are more core dependent than the Sandra scores.
So the memory bandwith of the P2 running L2 at the same speed as the P1 MMX is 93.1 MB/s to 128.7 MB/s. The P1 is about 38 % faster there, and 23 % faster at MMX. But if you look at the Whet values (177 / 153), the P2 is about 16 % faster. Assuming the MMX part is compareably fast as the the FPU part, this could just be the difference between the 23 % and 38 %.
The Int MMX test routine of Sandra seems to be overall quite memory speed dependent as the K6-III+ scores so unrealistic high.

Also the K5 scores are a bit strange for FPU. Its quite fast at Sandra FPU, even faster than a P1 MMX, but very slow at Whet. Though Quake performs very good.

K5 has very impressive ALU performance so maybe there's something to consider there.

You are right, this could it be. After all the typical part of FPU code is about 5 % and with highly optimized code maybe 15 %. (It doesn't say how much time the CPU needs to for it)

Hey Swaaye, GR100's topic about the 486 PCI divider gave me an idea.

I remember the FIC PA-2005 (socket 7 1997 AT board) has the ability to run with single SIMM's installed, running a 32 bit memory bus instead of the usual 64 bit.
What would be interesting is a comparison of a 486 board running at 133Mhz with the PCI divider enabled and the PA-2005 running a Pentium on a 32 bit memory bus. This would match the 486 and Pentium hardware in so much, it would be virtually as if they are on the same board, a comparison between the 486 and the Pentium as direct as it can possibly get 😉

Once I get my attic back again, I'm gonna start building me a couple rigs hehe 😁

You need to add downclocked Pentium II/III chips for comparison with the K6-III+. If they don't downclock that far, then maybe a new challenge that is suited to whatever their lowest downclock might be just to see how they compare directly with the K6-III+ because it really isn't fair to stick that one in with the rest of these chips that are older, slower designs.

Tetrium wrote:

Hey Swaaye, GR100's topic about the 486 PCI divider gave me an idea. […]
Show full quote

Hey Swaaye, GR100's topic about the 486 PCI divider gave me an idea.

I remember the FIC PA-2005 (socket 7 1997 AT board) has the ability to run with single SIMM's installed, running a 32 bit memory bus instead of the usual 64 bit.
What would be interesting is a comparison of a 486 board running at 133Mhz with the PCI divider enabled and the PA-2005 running a Pentium on a 32 bit memory bus. This would match the 486 and Pentium hardware in so much, it would be virtually as if they are on the same board, a comparison between the 486 and the Pentium as direct as it can possibly get 😉

Once I get my attic back again, I'm gonna start building me a couple rigs hehe 😁

Or you could just put a POD on a 486 motherboard. It won't hit 133mhz, but the chances of it hitting 100 are pretty good so you could do a 100mhz comparison of those chips.

sliderider wrote:

You need to add downclocked Pentium II/III chips for comparison with the K6-III+. If they don't downclock that far, then maybe a new challenge that is suited to whatever their lowest downclock might be just to see how they compare directly with the K6-III+ because it really isn't fair to stick that one in with the rest of these chips that are older, slower designs.

But Swaaye has already included the Pentium II. The unlocked pentium III E.S. that I have won't go below multiplier 3.0x, so that is not a fair comparison.

sliderider wrote:

Or you could just put a POD on a 486 motherboard. It won't hit 133mhz, but the chances of it hitting 100 are pretty good so you could do a 100mhz comparison of those chips.

Yes, but that would exclude many of the other chips again.
The only way I can think of to successfully overclock the Socket 3 POD would be to find a way to increase it's voltage, which would require modding of the CPU itself as it already runs at 3.3v with the motherboard set to 5v 🙁