do you have a rapidcad? i'm curious where that would fall between the 386 and 486

Oh the nostalgia. I did my PhD designing maths co-processor hardware. My rose-tinted glasses filter out the hideousness of having to design these things and now I can just appreciate the beauty in them.

I remember in 1993ish when the AutoCAD ppl at work who were using older computers got upgraded with 387 co-pros. It was a glorious day.

Interesting to see that Cyrix nailed the fpu performance in the 387 era but got beaten up two generations later.

These chips were the Radeon Instinct MI 100 and Nvidia Tesla A100 from the late 80s / early 90s. Amazing how much this tech has progressed in 3 decades.

Finally this channel gets the traction that it deserves. All this content is incredibly interesting, and I’m not at all into collecting stuff. It’s just a superb way of explaining the fundamentals of micro architecture. (edit: typo) (edit 2: and what in the world is that EDM-meets-yodeling song? Googling Austria hits, yodel techno, etc. is going nowhere kzfaq.info/get/bejne/p8d3hNuLqr_GqWQ.html )

I remember all those advertisements for CPU with math coprocessor.

who knew that I would need to wait 30-ish years to see benchmarks regarding something I always questioned myself about. Thank you !

Oh how I wanted a co-processor for my 386 SX25 all those years ago. Now I know that my desire was warranted 😄. Excellent video 👍✨

In early Pentium days, games started to require "486 or better", but they usually only needed a math coprocessor. By that time, 387s were very inexpensive and, most of the time, they were enough to allow 386s to run "486 or better" games.

Awesome! I Nudged you for this maybe two months ago, thanks for doing it! It's a historically valuable video.

There was also in the days a very unusual brand, Weitek, their coprocessors were the fastest, but they didn't work as a x87 replacement, needing a special socket. In fact they were available for other CPU families as well. Also kudos to all those manufacturers that took the effort to implement IEEE 754 in hardware, it's quite a complex standard for 1985's technology. (EDIT: Oops, I have realized someone there is at least another comment on the matter)

Hooray for Crystal Dreams! Loved that demo!

I had to get the 387 for AutoCAD 1.0 . The difference really was impressive for my old i386. Drawing in 3D was so very slow without it....painfully so.

I remember MicroProse Grand Prix (GP1) speeding up when I installed a 387 on my Cyrix 386SX-16 computer back in the day. Might be an option for your 387 testing.

The Co Processors were very much relic of their time, but theyre still AWESOME!

Your channel is very addictive 😍 Love your videos, thanks 😊

I remember contemplating getting a Fasmath back in the early 90s when they started to get really cheap. I saw benchmarks of one being ran at 16mhz on a 33mhz intel 386 DX and seeing it almost keep up with the intel 387 at 33mhz and was very very impressed. What I didn't understand at the time was the first few Cyrix FPUs were limited to lower clocks, but when ran half bus clock they were able to avoid async latency penalties and their IPC was increased a measurable amount relative to clock (vs say running them at 20 or 25mhz on a 33mhz bus clock). So it made them look really really good. Of course running them at bus clock will be much faster but the ability to show it off at 16mhz was a great selling gimmick. Of course the Intel FPU would increase IPC in the same situation, but the salesman wouldn't ever bother divulging that. IIRC halving the clock of the FPU on a 33mhz bus would only slow it to about 60 or 65% and not to %50 like one would expect. Perhaps due to a relatively higher amount of memory bandwidth. This is why there was usually a minimal difference when I ran mine at 20mhz over 16 mhz in my computer at the time. Faster clocks were always better, but some dividers had less penalties than others. I saw no reason to run mine at 16 other than to impress people, but being a little bit OCD it just seemed more rounded. I am unable to recall however if the slower FPU would end up bottle necking the 386 at all due to it potentially tying up the system bus a little longer? I didn't think it an issue at the time. Some integer ops ran slightly slower after putting in the FPU (like maybe 3-5%) but from what I understand that was normal no matter what speed the FPU was. For what I was doing, it was a welcome boost and well worth every penny.

Wonderful stuff!!!

the cyrix really kicked ass

I love vintage hardware! (And your channel)

Awesome for screensaver demos ... amazing stuff

Very nice dude ! You make my day. Now I understand that the ULSI FPU is a good choice (I just ordered one) Thank you for your nice job :)

Nice comparison!

Great video with very nice comparison. I like this channel because there are very interesting informations.

Love the video. I would also love to see these benchmarks run at the other CPU-s. If you can control for run-to-run variances, the fractal test could show even small differences, or maybe pinpoint some specific architectural traits.

Very interesting video. I used to own the IIT DLC3 387 math chip in my old 66Mhz IBM Blue Lightning computer which I purchased back in late 1993.. First time I've seen that chip in years... quite nostalgic.

I messed around with POVRAY back in the DOS 5 days on a superb 486SX25. Renders would take hours to days to finish. When I upgraded to a DX2-50 it would take minutes to hours to render. I never did own a socketed co-pro.

You really have to add the Weitek co-processors. They needed a different socket, but they were fast! I had done a similar (but more comprehensive) test about 18 years ago, comparing everything @ 20MHz, with what 387s I had available, but also with various 386 models. Apart from one graph (on some CPU forum which I cannot remember), I never published the results.

Another great video:)

These old floating point chips make me appreciate the FPUs that come standard in today's modern CPUs, especially given how expensive these chips were back in the day. As a regular home PC user back then, these chips were nothing more than curiosities.

Our first 386SX mentioned a Weitek co-processor

As a child I came across different systems which had those empty sockets for Math CoProcessors and I always wondered what magic could be in there. Now I know for certain! Thanks so much, I love this channel!

A great topic for those who don't know how we used to suffer back in the day! I appreciate this video and thank you so much for your hard work! Your brother, W.Bushnaq from Kingdom of Saudi Arabia

You have an impressive collection of hardware there! Respect. Edit , in 90-91' I was still running an Atari 1040 STe and an Amiga 500. The hardware you are showing was pretty high end/niche and indeed quite expensive at that time

Great video!

As an 386-enthousiast I can only like this kind of stuff. Very nice! Interesting and surprising results. I also have a selection of 387-FPU's in my collection; but I never bothered to actually test them like this. Just as you, I don't like empty sockets, so just installed one of them... I now learned that I did choose one of the slow one's ;-) (the ULSI MathCo).

if i remember correct, the Cyrix also has 4x4 Matrix functions build in, i don't know if any of the other brands has this feature, i think there were some libraries to bind while compiling.

Great test and comparison. I myself have in my 386DX 40mh, IIT 4C87 40Mhz coprocessor

"One of my personal favorite units."

Thanks for this!!!

Nice video as always :) I wanted to see this comparison. I was collecting 387s recently. I have 4 of them. The top models from Intel, ITT, ULSI, and Cyrix. Waiting for some 386 chips to test these. My 386 collection is very poor.

have my old Am386-40 with an IIT 387 running a manufacturing machine. Still running strong every day after all these years, even with the old "cheap" vga monitor.

I remember a long time ago (mid 90s) I stumbled across a free 287-10and man it made a huge difference with Links 286. Iirc it was actually 2MHz slower than the main cpu, but they seemed to get along well enough.

Fascinating comparison, I remember there was a DOS utility that would emulate a coprocessor for CPUs that didn't have it, I wish I could remember the name of it as it actually gave a slight boost to performance. I recall wanting to buy a 387 back when I had a 486SLC and then being flabbergasted at the price - far outside of my range back then as a student!

I saw a familiar name in the list of FractInt contributors. Ken Shirriff.

This reminds of of when I started in IT, I remember so many of these things.

The only thing I found so far is that the Cyrix FasMath CX-83D87-33-GP (and KN) models are asynchronous unlike others so they shut down unused parts of the FPU to save memory bandwidth on the data bus. Otherwise they were made same, same 32 bit bandwidth and data bus and same frequency. Will update as I find out more information.

Thank you for video, interesting

Was wondering if it makes sense to add a copro to my 486DLC40 CPU, as it is annoying to see an empty socket as you say... At the end, it's only an esthetics matter, without any purpose. Ahhh, the best part of the best DOS demo, Crystal Dreams II! You gave good tastes 😉 Schöne Grüße aus der Steiermark!

Even there is a demo, nice! I had couple years ago Cyrix FPUs, now only chips which I left are old 68040 cpus.

This comparison would have been fantastic for me in 1994. Our primary family PC at the time had a 100MHz IBM Blue-lightning CPU and a 33MHz Cyrix Fasmath. I had always felt the Fasmath let us down for performance, but now I know it wasn't actually so bad and just the low clock speed that was an issue. The computer was upgraded with an Intel DX2 Overdrive which was noticeably slower for integer but much faster for floating point, and then later with an IBM 5x86 that was much faster in every way.

If I remember correctly worms was a game that was unplayable slow on a 386-sx 20... But got a huge boost with a 387. Would be cool if you could test that :)

This was a great video! I love fractals. I want to run some of these programs on my 10700k RTX2070 just to see how far along we've come.

Fractint ftw! One of my faves when I was a kid.

My recollection is that the 486SX processor was a full 486DX processor with it's floating point unit artificially disabled. The 487SX "math co-processor" was in reality another fully functional 486DX CPU, with one extra pin to prevent it from being used in regular 486 boards, and also to prevent a regular 486DX from being used in a 486SX board. Once you installed the "487SX", it took over completely from the 486SX which simply sat idle in a halt state. The only upside to this whole ridiculous arrangement is that it later became possible to install more powerful "486" and "586" processors with clock-multipliers to get a genuine performance boost.

I found your channel to be very interesting

Great benchmark ! It will be interesting to add the Rapidcad if possible !

This old stock looks fresh new with the quality of your prod! Glad I found you few months ago :)

I remember the Motorola 68881 math co-processor

Cyrix: Best FPU when it didn't matter, worst FPU when it totally mattered.

Would have loved to see a software emulator compared as well, man I spent some days in fractint. Great vid man, really enjoyed it.

Very interesting and useful information. Expected the IIT to score a little better than it did, for no particular reason, though Cyrix was always king in the FPU department back then. It's a shame they couldn't keep up in later years. Going to have to re-test my own Cyrix 387DX, purely as I have to wonder if they were the same as the 83D87 versions or not, but don't have one of those (or the mysterious KN version). It should now be possible to deduce this from looking at your results, figuring out how far off my testbed system is with known chips (Intel, IIT) and applying that 'offset' to whatever the Cyrix chip scores. So thanks for doing these tests, very informative.

im re-watching the video and im late to receive my ULSI DX/DLC 40 Copro to push my 386DX 20 over its limits! thanks again for your amazing contend , take care F

Seeing that 487SX I'm officially envious now. :) That's a damn rare find! I'm waiting for that separate video about it. Edit: Fractint was a great program. For hi-res pictures with more complicated math it could took a night to render. I still have a lot of those fractal images on CD. And some of them printed in color and hanged on the wall. :)

Very interesting fact on Simcity 2000 using the old FPUs!

I'm glad I found this channel. Love em processors. Also, I have a question: When this person was installing the co-processor, how did he manage to place it in the correct way? I didn't see any marks or guides.

It's amazing how Cyrix performed well with the KN and some years later failed in the same subject with Cyrix 6x86

I had one of those IIT FPUs. It could do 4x4 matrix multiplication. Good for graphics. Unfortunately it was not used by any commercial software I got.

I remember the day I installed my 387 copro, aaah those good all days....... !!!!!

If I remember correctly, that i487SX is basically a full 486DX that checks that a 486SX is present, then disables it and takes over all processing by itself. Seemed like a bit of a scam, especially if your 486SX was just a DX with it's FPU disabled. Looking forward to the video on that one.

I just bought a ulsi chip from china, and it was real! A perfect match for my ti486dlc. Pretty cool

I was messing around with some old computers a couple years ago and i discovered that some of the more recently compiled open source DOS tools were compiled in such a way that they require floating point. That was frustrating as one of the machines i was using was a passive backplane board with no accommodation for a 387 even if i had one. I recalled that back in the day there were fpu emulator TSRs for dos, and went looking, and found that you can still license them for a bunch of money. Oh well.

We had 8087 in XT-machine, then 387sx when moving to 386sx machine, after which the FPU was built in main CPU. My father used it for CAD-stuff, I merely for fractal rendering and Falcon 3.0. Funny how for me it was ”normal” to have FPU, but none of my friends had those. We then sometimes compared how long time it took to render some images, and with FPU it was a lot faster with proper software.

Interesting how they had identical scores, usually there's a little error marginal even when benchmarking many times with the same component :)

Excellent video! And of course now all the scalpers are going to start buying up Cyrix 33GP FPU's to resell on eBay for thousands lol ;)

i have a i9 9700k and i don't know what the big caps and 2 small caps or whatever on the bottom are called and don't know wher eto buy em. Can i fix a damaged pad?

Glad I've been holding off for months on deciding what FPU to buy. I really want my 386 (on one of those FOREX boards w/ clear jumper instructions on the PCB) to look proper...plus I'm trying to make it a bit faster than it was back when I last used it (around 2006-2007). Unless I find a 486 DX/2 66 cheaper.

about games using the math co, there is the original quake and a game called retro city rampage 486, they both run on 386 hardware. Also a lot of software like some trackers makes use of it.

I remember using some shareware that emulated the 387 co processor in software, for the programs that absolutely needed the maths unit in there. While it was not as fast as the actual silicon, it came on on most testing as slightly faster than the actual CPU itself in doing this, more from optimised code for the functions than anything else, so would probably have cut the CPU only render times down to under 40 minutes.

Yay, I had the ITT 3c87, for POVRay! (and demos + own coding ofc). It had a good price/performance ratio. I even called the c't magazine's hotline to help with my decision (I don't remember if Andreas Stiller was on the other end, but this is likely).

This example of FPU, is like also saying the comparison with Hypertheading when it arrived in CPUs.

Most interesting bit was the mother board which could fit 486 and 386 chips(I did not know it was possible), could you please repeat the same benchmark with 486 CPU?

Thanks for telling me i don't need one of those, I was just about to order one to try and speed up my Ryzen system.

Cyirx? Excellent Video.

As far as I'm aware the 487 is just a 486DX that takes over all functions from the CPU. There are some early 386 boards which allowed you to use 287 coprocessors, it would be very interesting to see these in action and how they compare to the 387 setups! Would also love a follow up to see how much better the 40MHz parts perform at their rated speed with the AMD CPU. Makes me wonder what the top dog best setup is.

Good vid. Really great content from your channel. My only grumble would be out of the box thinking. Ie could you have stuck a 486dx 33 comparison at end. Ie the generation leap to 486 with integrated fpu. Edit typo

Can the 40's beat the 33GP-KN if you pair them with the 40MHz - guessing there is no option to run the FPU at a different clock to the CPU

Do you have a 386 board with two FPU Sockets for Weitek and 387?

I'm curious about that motherboard you have used, would a 486 to Pentium Overdrive or equivalent chip work on that motherboard also? Yet another really interesting, detailed and well explained video. Great work. 👍🏻😃

I am curious, are there any differences between them other then performance? Are there any known compatibility issues?

We take floating point arithmetic in hardware for granted. My dad used to write these things in assembly language. In the second volume of Knuth's book "The Art of Computer Programming" there is an entire chapter on multiplication "How Fast Can We Multiply?"

One thing of interest might be how well the various FPUs complied with the IEEE 754 standard. There were some interesting "bugs" in various cores, and also in the emulation libraries. (For some values of interesting, of course.)

My father worked with CAD and other structural engineering software. I always inherited his computers. Dual socket 4x86 dx2 and dual Pentium pro 200 were the most insane ones. I would really enjoy a review of those systems if at all possible, as they are lost to time. :(

How do you know which way round the CPU/FPU goes in the socket. It can fit in the socket any way. If you get it wrong, the board and the CPU/FPU would be damaged.

Nice video. Does a co-processor work faster/slower with different main processors (brand, clockspeed)? Will it change the ranking? How much faster is an integrated math processor compared to a seperate co-processor? To be continued?

"But why are we collectors populating empty 387 sockets? Uh, you know the answer: yes, because we can and just hate to see empty cpu sockets." Made me laugh hard. But it's actually true. Even 287s if that matters. I have a few fpus and wanted to test them but decided not to because it would have taken time (sloppy excuse). After a few days I get this. Come on, coincidence in the galaxy? :)

Hi Sir, nice video! Can you please recommend an 387 for my 386DX40 build? It is only to populate the mobo. Do I need an xtal to go along? I have a xtal socket placed next to the 387 socket on the mobo. Many thanks!

Finally Cyrix won!! Wieitek was missing though

Mate, why not find the original Intel tool from the overdrive kits? I have one, it's perfect for the job of removing chips.