Anonymous Coward wrote on 2020-02-05, 01:55:

mpe wrote on 2020-02-04, 13:28:

But it wasn't that important as the 75 MHz was the only CPU to use the 50 MHz bus so IMHO it wasn't worth it.

This may not be *entirely* true. As far as I recall there was a second (far less common) CPU from intel that used the 50MHz bus...a gimped version of the P100 that ran at 2x50 (probably a substandard part). I'll have to check CPU World again for the S-spec, but I think this CPU might have also run at a higher non-standard voltage too.

Uuuuh .. A Pentium variant i don't have in my collection .. 😁

H3nrik V! wrote on 2020-02-05, 09:14:

Anonymous Coward wrote on 2020-02-05, 01:55:

This may not be *entirely* true. As far as I recall there was a second (far less common) CPU from intel that used the 50MHz bus.

Uuuuh .. A Pentium variant i don't have in my collection .. 😁

CPU world lists only Q0674 as the exclusively 50 MHz bus 100 MHz chip. Good luck finding one ! 😀

http://www.cpu-world.com/sspec/Q0/Q0674.html

On the other hand probably all 100 MHz chips are allowed to be ran at either 50 or 66 MHz.

mpe wrote on 2020-02-05, 08:50:

Not just the 430NX. But all Intel's chipsets up to 430TX drive PCI @25 with Pentium 75 due to the fixed 1/2 divider. Should anyo […]
Show full quote

konc wrote on 2020-02-05, 07:59:

So this means that on the 430NX the PCI does run at 25MHz? Really weird for Intel to have done this. I'm curious to see pure CPU benchmarks as well, something that does not rely on the graphics card/PCI bus too much.

Not just the 430NX. But all Intel's chipsets up to 430TX drive PCI @25 with Pentium 75 due to the fixed 1/2 divider. Should anyone know about a motherboard that can override this please share it. All my Socket 5/7 boards all have dividers for ISA bus only.

I think the PCI doesn't hurt performance that much. The cache/mem latency and 50 MHz bus is more important factor IMHO for frame rates in Quake.

Pure CPU benchmarks (like Dhrystone) are obviously proportional to CPU clock - chipset has no effect here.

The PCI bus really does hurt performance in tasks that require bandwidth - do a Quake benchmark and you'll see the P75 really suffers against the P66.

Too bad.
Oh well, the additional benchmark I was hoping for can probably be done on another board and scaled accordingly. Thanks for the feedback! 😀

The Quake test is almost exclusively CPU (FPU/cache/mem) bound load when testing CPUs of this gen. Even at 320x200 those CPUs struggle to draw enough data for the bus speed to matter.

I was looking into this when comparing 50 MHz PCI and 25 MHz PCI when overclocking my Am5x86. There was much less difference than I thought. PCI isn't that sensitive to clock (as ISA or VL-Bus is for example).

Perhaps the DOOM is a little more bus bound, but still I don't believe 25 MHz PCI is a big bottleneck.

mpe wrote on 2020-02-05, 09:21:

CPU world lists only Q0674 as the exclusively 50 MHz bus 100 MHz chip. Good luck finding one ! :) […]
Show full quote

H3nrik V! wrote on 2020-02-05, 09:14:

Anonymous Coward wrote on 2020-02-05, 01:55:

This may not be *entirely* true. As far as I recall there was a second (far less common) CPU from intel that used the 50MHz bus.

Uuuuh .. A Pentium variant i don't have in my collection .. 😁

CPU world lists only Q0674 as the exclusively 50 MHz bus 100 MHz chip. Good luck finding one ! 😀

http://www.cpu-world.com/sspec/Q0/Q0674.html

On the other hand probably all 100 MHz chips are allowed to be ran at either 50 or 66 MHz.

Oh, well - I've decided not to collect ES's unless I directly stumble across them .. I hoped it was like the 486, where the SK051 spec of DX4-100 actually is capable of 2x50. Well, I'll stop being OffTopic now ..

mpe wrote on 2020-02-05, 09:41:

The Quake test is almost exclusively CPU (FPU/cache/mem) bound load when testing CPUs of this gen. Even at 320x200 those CPUs struggle to draw enough data for the bus speed to matter.

I was looking into this when comparing 50 MHz PCI and 25 MHz PCI when overclocking my Am5x86. There was much less difference than I thought. PCI isn't that sensitive to clock (as ISA or VL-Bus is for example).

Perhaps the DOOM is a little more bus bound, but still I don't believe 25 MHz PCI is a big bottleneck.

This is not my experience, for both Quake and Doom.. The P75 runs like a dog in both. If I were to ballpark it I would say it is about on par with a P60.

You can think about this this way:

Both Doom and Quake use sw rendering and 320x200/256 resolution (although former is Mode-Y and the later VESA, but that's just a technicality). So if engines were equal they would be transferring the same amount of raw graphics data over the PCI to the VGA to render a frame.

Obviously that's not the case as the Quake engine is much more complex (CPU bound). Yet the Doom is able to push twice amount of pixels to the VGA without being capped by the PCI.

You can also prove it experimentally by putting in a 150 MHz CPU instead of 75 MHz (or other 200 instead of 100 for that matter) which almost doubles the framerate in Quake. Despite sitting on the same PCI bus. That wouldn't be the case if there was a bus bottleneck.

You'd need a way faster CPU to saturate the PCI bus by Quake. Like a magnitude faster. Even at low res. By the time that happened games stopped being rendered by CPU (and also new bus standards emerged).

I did not believe you so I went ahead and looked at the Quake results on thandor.net.. https://thandor.net/benchmark/33

And yeah, you are right, the P75's result falls right in line with where you would extrapolate for a 60MHz FSB processor.

appiah4 wrote on 2020-02-05, 10:56:

I did not believe you so I went ahead and looked at the Quake results on thandor.net.. https://thandor.net/benchmark/33

And yeah, you are right, the P75's result falls right in line with where you would extrapolate for a 60MHz FSB processor.

...except that P75 is benchmarked on an i430VX board with PLB cache and either EDO or SDRAM (not sure which Thandor used), where the P60 was running on an i430LX board with asynch cache and FP. That difference alone should be good for at least 15% performance, so in fact being on exactly the same line with those platforms means something is severely impacting performance.

Question is whether it's the slower PCI bottlenecking the P75 or whether the 50MHz memory (vs 60MHz) is more relevant. Given the 15% delta vs expectations, which is more than just RAM clock, I'd say it's probably both.

Woot, I am vindicated!

mpe wrote on 2020-02-05, 00:47:

It basically confirmed my expectation. The Pentium 66 is quite a bit faster that the P75 when compared on a like chipset. Clearly the bus speed does matter when other things are comparable.

What I see in these results:

The P66 system is about 8% faster at the maximum, not what I personally would call "quite a bit". But yes, that was to be expected, since you only did benchmarks that heavily rely on graphics. Of course the system with the higher bus speed and hence faster access to video memory wins. I fail to see what this test should show..? You have the P75 with only 12% more processing power, but the P66 with 25% higher bus speed. Had you only chosen CPU-intense benchmarks, all results were inverted with the P75 being the winner. And that's the problem with such tests: You subconsciously select benchmarks that will confirm your expectations in the first place.

mpe wrote on 2020-02-05, 10:37:

You can also prove it experimentally by putting in a 150 MHz CPU instead of 75 MHz (or other 200 instead of 100 for that matter) which almost doubles the framerate in Quake. Despite sitting on the same PCI bus. That wouldn't be the case if there was a bus bottleneck.

That's not really how it works. No saturation of the bus is needed to make it a bottleneck. The slower the bus is, the longer it takes for the CPU to write a rendered frame to video memory. This time taken is no longer available for actual calculations. So with the faster bus, the P66 simply has more time left for calculations than the P75 that is wasting quite a bit more time for writing the frames to video memory. See above for numbers: the P75 can only calculate 12% faster, but the P66 can push data over the bus 25% faster.

I would also suggest bench mpeg encoding for example with a test snippet, superPI and such Stuff.

i also have both boards

Or benchmark games that rely less on graphics but processing power, like Chess games, turn-based stragety, city simulators. Problem is, these normally don't have benchmark functions. Well, Chess games maybe, not sure.

Actually if I was prepared to put my money where my mouth was, I could do this 'level playing field' comparison too - I have an Asus SiS501-based So4 board and someone is locally selling its So5 twin (together with a nice P3C-E and some relatively uninteresting cards). Unfortunately he isn't prepared to split the lot and even though the price is reasonable, apart from this curiosity I have no valid reason to buy either and too many other things to spend money on, so too bad I'm afraid.

But tell him to do it, for science!

This is such quality content with geat analysis and discussion. The evolution of the original Pentium platform is fascinating to me. There are so many changes from 1993 through 1997: three sockets, multiple voltage changes, six unique intel chipsets, three process nodes, an over 3.5x increase in clock speed, and so much more. It's also the time of transition from DOS to Windows 9X being the predominant PC operating system. It's such a cool time in computing history.

derSammler wrote on 2020-02-05, 14:52:

The slower the bus is, the longer it takes for the CPU to write a rendered frame to video memory. This time taken is no longer available for actual calculations

Perhaps I should have phrased it better. It is not that the PCI bus speed is completely unimportant. My point is that the Quake runs at relatively slow framerates on this gen of CPUs (about 20fps) even at 320x200. That’s primarily because of the complexity of the game engine as there are simpler games which runs can run much faster. If you think about a profile of its game loop, where it has to deal with relatively complex geometry, update AI, do all the texturing and render the frame and other stuff, then the amount of time spent on retiring the relatively small frame buffer into the video ram is relatively insignificant. The final blitting is also heavily optimised (not every pixel is transferred every frame), PCI writes are also not always blocking (write post buffers and bursts). So even if you do that part in theory up to 30% faster (33 vs 25 MHz PCI) it is just just a fraction of fraction to make a difference.

I should have a working Quake for DOS source compiler toolchain in one of my VMs. When I have time perhaps it would be interesting to see if I can disable the final image blitting or render to a off-screen memory to isolate the PCI performance. That would make also a nice non-graphic benchmark as some suggested...

Great comparison!

I overclocked the P60 I had to 75mhz and will scan for the screenshots that I took. It was not really fast though. 80mhz did only work for a short amount of time - the northbridge needed cooling from 75 upwards.

My point is that the Quake runs at relatively slow framerates on this gen of CPUs (about 20fps) even at 320x200.

About 5% of what PCI video card theoretically can achieve with much more powerful CPU.