FPU performance boost

[Tolgod]

Hi there.

I have a question in regard to the boost an fpu might give to a program.
The main reason I am asking is because I use Ibrowse and see that it supports one. I am considering to upgrade my 020/40mhz, but want to know weither I should consider the fpu a must.
So in real life performance, does anyone have any idea in what order of boost I should think.
Will it really unload the CPU and make it all much faster, or is it more of a slight performance boost on specific tasks.
In other words, will the non fpu and fpu version really feel completely different, or is it more a slight edge the fpu might give over no fpu.
I understand these type of questions are tough to answer, there are many factors, but I'm hoping for some real life experience maybe so I can decide weither it's a must have or just a nice lil extra.
Tnx in advance

[Kin Hell]

It's a bit like this fella....

A CPU without an FPU can not do Floating Point Instructions.

FPU's can accurately compute addition, subtraction, multiplication, division, and square roots.

They are specifically designed to compute Floating Point Numbers & off-load the CPU from having to do these Intensive calculations.

So yeah, if a Program can use one, it will be significantly faster than a CPU without one.

Try Ray Tracing an Image without an FPU & see how much longer it takes.

[Tolgod]

Tnx and that indeed seems logic. My issue remains though that I have no clue as to what significant means. That it will improve seems obvious, but what would I notice in real life. Networking is CPU intensive and browsing as well , so every clock cycle sopared will improve speed. I can imagine for a program like Ibrowse it will be used to render the images for instance. What I am wondering is by example lemonamiga.com. The home page is loaded in 98 seconds at 40mhz. If my 020 had an fpu, would it double the speed or even more, or might it give a 10 or 20 percent boost.
Again, I understand that there is no easy thumb rule for this as it will very much depend on the software and how much fpp is going on, so I'm hoping someone here might have some real life experience with this type of use and the actual influence an fpu will have on it.

[Kin Hell]

Have a Read here fella.

https://dl.acm.org/doi/pdf/10.1145/1499310.1499325

What Amiga are you running this on & what CPU card?

[Tolgod]

Tnx again. It's an interesting read, but doesn't answer my questions as there is only data on the fpu performance and I can't compare. Furthermore I still don't have a clue in relation to a program as ibrowse as to how much of the processing required is floating point. So the real life impact is what I am after. I'm running an a500+ with ACA500+ and 1221lc. Since my psu doesn't handle the 40mhz overclock very well, I currently run it @26.67mhz. There is 2mb chip and a lil over 9mb fast ram.

[Fastdruid]

Originally Posted by Tolgod

Again, I understand that there is no easy thumb rule for this as it will very much depend on the software and how much fpp is going on, so I'm hoping someone here might have some real life experience with this type of use and the actual influence an fpu will have on it.

If you have a suitable machine to run WinUAE this seems to me the ideal scenario to use it.

Setup a profile with appropriate hardware and software (if on CF you could easily clone your existing drive) but no FPU and time how long it takes to load a page. Ensure cycle accurate and not JIT is set as you're not actually interested in running as fast as possible.

Then do the same again but with an FPU.

It's not going to be directly applicable as the network stack will be somewhat different and may be hard to tell if there is only a 1% difference but it would be obvious if you see for example a 50% improvement.

[meynaf]

Originally Posted by Tolgod

Will it really unload the CPU and make it all much faster, or is it more of a slight performance boost on specific tasks.
In other words, will the non fpu and fpu version really feel completely different, or is it more a slight edge the fpu might give over no fpu.

The FPU isn't something that offloads the CPU for its routine tasks, it's something that handles floating-point operations in hardware.

So the amount of performance boost hugely depends on how much the target program relies on them. In addition, the program must explicitly use them and coders are often reluctant in doing so because you either have two versions of the same code (one with and one without), or a program that will crash on non-fpu machines.

For Ibrowse i'd bet it's javascript that takes most of the toll and that merely loading a page and displaying it, even with images, remains mostly unaffected.
Thus if it's about page loading, better drop the idea of an fpu. A faster cpu would have a much greater impact.

[Tolgod]

Hi all and tnx for the wonderful input. The emulation idea seems like a good suggestion and just to scratch the itch I'll be testing it. The last response is kinda what I was asking for, it seems like the balance would swing to slight performance boost rather than huge. I see in my own setup that there's an almost linear relation there with cpu speed and performance. On the other hand it does seem that the sites with more java script suffer the most (aminet loads under 10s), so the emulation setup is definitely worth a try. Tnx all, I'll report back the emulated data asap.

[Mathesar]

Originally Posted by meynaf

The FPU isn't something that offloads the CPU for its routine tasks, it's something that handles floating-point operations in hardware.

So the amount of performance boost hugely depends on how much the target program relies on them. In addition, the program must explicitly use them and coders are often reluctant in doing so because you either have two versions of the same code (one with and one without), or a program that will crash on non-fpu machines.

For Ibrowse i'd bet it's javascript that takes most of the toll and that merely loading a page and displaying it, even with images, remains mostly unaffected.
Thus if it's about page loading, better drop the idea of an fpu. A faster cpu would have a much greater impact.

I always disable javascript when coming here
Wouldn't an FPU help with faster jpeg decoding?

[meynaf]

Originally Posted by Mathesar

Wouldn't an FPU help with faster jpeg decoding?

Everything in jpeg decoding can be done with pure integer without any speed penalty. Actually, might even be faster this way.

[Tolgod]

Hi there,

As promised I would come back with some emulated data.
Now I've spent all weekend litterally doing litteally hundreds of measurements, so hopefully someone else at some point might find the info usefull.
If I learn one thing it is that I have a lot to learn about winuae.
It does however feel good to have some general working understanding of it now and I will find a new hobby in all winuae. The real deal will however always remain my weapon of choice and is uncomparable in emotional impact to the emulator.
I have done test on two different systems, one p4/winxp/1GB and a C2D/win10/4gb.
Now I'm presuming that different architectures will have a specific setup (os and winuae version) that will be ideal, but I can also absorp so much new knowledge in a weekend, so the pefect setup will be a long term quest from here on.
Across the board I did get similar readings thoug.
All are loading time of lemonamiga.com.
The general trend that is shown is that an FPU has little to no influence, other than less stable operation and guru error (yellow and red) on occasion with Ibrowse.
Also a 030 at equal speed was in all setups slower than the 020 funny enough.
Having installed Ibrowse a couple of ties over the weekend I saw omething I hadn't noticed before, the installation only has four options in regards to architecture :
020
020 fpu
040/fou
060/fpu

In other words, no specific 030 option

The results measured are as follows :

P4/winxp/1gb

020FPU/28.x mhz - app. 1:15
030FPU/28.x mhz - app. 1:28
020/28.x mhz - app. 1:15
030/28.x mhz - app 1:35

C2D/win10/4gb

020FPU/40.x mhz - app. 0:51
030FPU/40.x mhz - app. 1:02
020/40.x mhz - app. 0:47
030/40.x mhz - app. 0:56
040/40.x mhz - app 0:32
020/28.x mhz - app. 1:02
030/28.x mhz - app. 1:17
040/28.x mhz - app. 0:34

Now all noted down times had to keep reoccuring and they did, at a point it became predictable even.
So an fpu won't be a magic bullet for Ibrowse, and the performance loss on 030 seems to be yet another sign that Ibrowse and it's ridiculous (though no inflation was charged :S (in a completely deflated market) 55eu pricepoint is a travesty.
Choked to death for 12 years, hardly updated for 68k, most of the work went to os4, no price differentation for new users (where updaters can choose 68k only), and then just slapping the price of twelve years ago on it. So I would be paying for the development of a platform I'm not using.
This chokes our community, so much more people would be lured in at a pricepoint between 20 and 30 and I'm 100% convinced it would earn more, and then there would be a much larger customer base for updates.
So it's ecinomical madness, pure greed and no love for the community as so many goodwilling amigans gracefully reason.
It's choking the community at this price point, it will lead to eventual death as the base will remain too small but more importantly, there are many people developping and supporting the community for much more reasonable prices.
Fully new developped boxed special edition games go for less, Roadshow from 25eu, os 3.1.4 from 25eu etc.etc.
I happily spend hundreds of euro's on my amiga and also pay for software, I do have to have a feeling though that it is a fair deal and the recipients of that money have earned it.
Anyways I'm straying of into a rant and things are what they are.

For now I think I going to wait out the development of some 040 cards that are coming and I will get the powersupply from Icomp once released.
The not knowing wheither all is powered well can't go on, I need to know all is good and safe on that side.

Tnx all here for thinking along and hopefully someone will get some use out of this crazy exercise I've been doing over the weekend.
Grtz y'all

[meynaf]

Timings on emulators for anything above 68000 are unreliable. Do not draw conclusions on them.

[Toni Wilen]

Emulated FPU isn't even meant to be timing accurate. Also there are 3 different FPU modes. 64-bit is usually much faster than 80-bit (almost no program cares). Softfloat is much slower but bit perfect accurate.

Real world 6888x are quite slow anyway, most FPU instructions take at least 50 cycles. 68040+ FPU is much faster.

[Tolgod]

Well, I do see a vast difference with real life performance, but the trends are the same across the board.
The emulation suggestion was done here, and crude as may be, it's the only viable source of some more info in regards to the questions I'm having.
Now I'm not using any diagnostic tools or anything, just lookin at how Ibrowse performs on loading a certain website.
Unless someone here actually has both setups and could provide me with real life data, this is the best hint of what to expect.
The conclusion I find safe to draw is that an FPU won't be a magic bullet for browsing and an upgrade to a 030 system neither.
That in real life the 030 might outperform the 020 on the use of IBrowse is still possible, though the results on emulated setups suggest there that IBrowse isn't particularly helped by the extra computing power of the 030.
What I do find safe to assume from this is that a 030 won't likely give me more of an edge than that which is measured on real systems in the sense of drystones per clockcycle. In other words, it might be quicker, but also no paradigm shifting performance boost. Currently my 020@26.67mhz loads lemonamiga in app. 1:41.
I now doubt a 030 at equal speed with an fpu could do that in under a minute.

Let me just put out the request explicitely to anyone reading this and having a real life 020 and 030 setup and networking solution, could you try to load lemonamiga.com in Ibrowse on both setups and report me the difference here?

There, I can't do more than that, for now this all I have to go on.
grtz

[meynaf]

If all you want is to load pages faster in ibrowse then you can achieve this goal without changing anything in your hardware. Just setup ibrowse to disable image loading.

[Tolgod]

Well. That is obviously something I did try early on and didn't seem to make much difference. At the time though it was just the general experience, nothing was clocked. So I just revisited the notion out of curiosity and in the case of lemonamiga.com it cuts the time in half going to 51 seconds rather than 91.
Still nowhere near 'browsable' (which I consider to be at least under 30s) but indeed much faster and on a double clocked system perhaps even ,browsable'.
In all honesty though, suggesting this workaround for the issue is kinda like suggesting total abstinence as the best solution for birth control. In it's self not untrue, but it leaves much to be desired.

[meynaf]

If you find some image is of relevant importance, then you can still load it separately. It is not a pure loss.

[Tolgod]

That's true, I can always 'marry' the image :'D . Won't know how I like it though until I do ;p

But it is a fair tradeoff for some more modern functional sites like some fora or a price comparisment website I often use. Those ones won't load at all it seems, but without image can function, albeit still slow. I do however wanna explore how I could get some of the simpler mostly amiga, retro game/prog or emu related websites working at better speeds with eye candy on. I'm already pleasantly surprised with how it functions now. But I'm still considering an upgrade and in order to decide weither I'm upgrading or not and how to upgrade if so, the impact on this use of my beloved Amiga is a factor strongly weighing in.

[AmigaHope]

An FPU will not decode your JPEG images faster than a fast integer decoder, but it will decode them more accurately. If you use an *accurate* integer decoder, an FPU will decode them faster.

The 680x0-series FPU (68851/68852 or built-in ones on in 040/060) is incredibly accurate (80-bit IEEE floats), and simulating that accuracy in an integer CPU is very slow.

So it depends on if the code you are trying to run does fast-and-dirty JPEG decoding in the integer version.

[nikosidis]

Any browser for 030, 040, FPU or not, will be to slow. At least turn off image loading even if that is kind of annoying. There are many things. Slow internet speed, slow disk loading, low res. etc. Aminet works fine, that is the only web-page I use with my A1200/030. I use FTP so I can go into my Linux laptop and download files. Regarding the topic, FPU can boost programs that are written to use it. There are a few programs that come with own FPU version. There are also some demoscene productions that require FPU.