Man, as a younger cat who first got into computers around 2010, seeing all of these old school CPUs with ZERO cooling solutions literally being removed with a miniature crowbar is otherworldly! haha i love this channel, so rich in PC history!
Thank you to your subscriber who lent you the chip. Now we can have our cake and eat it. By the way having the "487" and the weitek at the same time was a nice fix. And one of the rarest setups of all time.
I always wondered what the Weitek option was in the BIOS of old 386 machines! Thank you for explaining it. I was also curious how the performance of those benchmarks are on 486 and Pentium machines. Had intel already passed Weitek with their newer chips?
Yeah, with clock double technology on 486 machines Weitek was out of the game and the Intel FPU was faster. I am still struggling to get a 4167 Weitek to proof that.
I remember the days when these processors were for sale. I think they had some double or triple the price of a 387. And the 387 was not cheap either. Both chips were absolutely necessary for CAD, as it took hours to turn a complicated wire model of a cathedral 180 degrees without a coprocessor.
Awesome video. As a kid growing up with Weitek in the BIOS this explains a lot. Still, a Am386 DX40 with Cyrix FasMath 387 would put that Weitek to shame. History is fascinating.
Love the new video intro. Intriguing set of chips and quite attractive. I would love to see it working in a CAD programme. Must have been the hot-shot chip to have in those days. 👍😊
This makes you wonder if there would’ve been any software available to take advantage of two FPU’s? Because as demonstrated, this setup is capable of parallel processing - in fact, you could say it is a 387 times three.
ruclips.net/video/nytpyJhDOHo/видео.html Quake demo on a Cyrix Cx487s is about 1fps, using both FPU's in this case would give outstanding 3 fps ! Well ... I definitely wouldn't play it, but I'd love to see it as a prove of concept. ;P
You could have tested the 67fonge and the 87fonge in the Rapidcad+Weitek configuration in one go, without swapping chips. Also it is always a bit suspicious when the test and comparison software is from the company who wants his product to be the superior, too bad you could not find a 3rd party test program.
I love this channel! You got me with the box opening though, I thought maybe you were about to do an LGR style opening and cause all sorts of collector rage :)
Thanks for the memories. I have a Compaq 386/20 and installed a daughterboard with a 387 and Weitek on it. I seem to recall around a 5x speed improvement for a chain of multiply/divide instructions. I did not have any compatible software for it so I wrote some test programs in assembly. One was a Mandelbrot generator. It was about 4 to 5x the speed of the 387. I didn't go any further with practical uses. Just a hobbyist around 1986.
Maybe you'll have the time to write a sample test for double FP precision for comparison with 387. ;-) I'd never get one of these boards but it did not came with the 3167 package. It was more an add-on to support both FP accelerators at once for special purpose. Did you have the company name by hand?
@@opensparcbox I think I did those kinds of tests back in the day. The claimed 4x to 5x speedup was confirmed. I was able to achieve 90% of the theoretical throughput in my little Mandelbrot generator. The 387 did 80-bit double-extended precision internally. The Weitek only did add, subtract, multiply, divide, and square root. Not transcendental functions such as SIN, COS, TAN, ATAN. So to get full floating-point support I'd have to write all those libraries and that wasn't something I was interested in.
@@billr3053 Thanks. I appreciate your answer. The 387 do 80-bit precision internally for all computations. So that should be near to equal, independent on the data type used (float, double). The 3167 on the other side using 31 float regs, each 32Bit wide which could be paired up to 15 double regs 64Bit wide (edit: one always has value zero, so only 15 double prec reg pairs) but then with computational performance degration, by just using supported instructions in hardware. So I'm wondering: a 4x to 5x speed improvement seems to be realistic for float type based computations only. Do you still have the assembly sources from this Mandelbrot generator? Would be nice take a look. :-)
@@opensparcbox Yes the 'single precision' was faster than 'double' on Weitek. As for the sources, hmmm. I can fire up the ol' Compaq and hopefully find those including the mapping and macro stuff I did to make programming easier. I remember optimizing and the way multiplication was done with so-called complex numbers was very symmetrical with crossing lines to the various registers. All partial results were kept on-chip.
Awesome video! The Weitec FPU at only 25Mhz really surprised me on how much faster it really is compared to that Intel 25 MHz as well. I can only imagine how much faster that FPU chip would be if it could have been used at the 33 megahertz or even a 40 megahertz version was available at the same time Intel or AMD making a chip like Intel's rapid CAD that could have been at 40 megahertz instead of just the AMD 386 DX 40Mhz alone since that chip does not have a FPU in it. Even the Cyrix 386 to 486 66Mhz upgrade and a Weitec FPU running at 66mhz (would need a 133 MHz Crystal)would have been awesome. Sad that Weitec chip did not have better software support other then mostly CAD like software. That really could have really pushed the 386 even more(the ultimate 386) during the 486 days, but sadly it was not ment to be. Great video and thank you.
I noticed that SI identified the RapidCAD as a 386SX, though... so it's limited to 16MB of memory on a 16-bit path? Sucky. Then again, even CAD software back in the day rarely used more than 16MB of memory, unless you were trying to do a very complex design... but I would think that the memory bandwidth cut might be an issue with general performance. Hardly a "486DX"... more like a 386SX/387 combo on a single socket. As far as I know, math coprocessors in general had very little support back in the day, and the Weitek especially so. I can't think of much other than CAD, raytracing, or fractal generation, that supported a math coprocessor. Would have been nice to see some ACTUAL CAD running, or even do a run with POV-Ray (the raytracing software that you use if you're "pauvre" because it's free ;-) ). That could also have been run on a newer processor like a Pentium for reference (which I'm sure would blow it away, but just for perspective). I don't think he tried hard enough to "make some testings," and for that I give a thumbs down.
Same since PC-AT time, when one KB was switched between test bench and my main PC... no KB connected and please press F1. Good one was so AT KB may be plugged hot without powering OFF PC. PS2 Mouse always required restart.
This video explains something that puzzled me many years ago! I was at a computer fair in the early 90's and I saw one of the sellers had some random motherboards one of which looked like it had 2 processor sockets on ir... I asked the seller about it, and he said "no, its for the maths chip" which at the time I thought was strange because to my knowledge at the time coprocessor sockets were smaller than the processor socket. So probably this board was one for the Weitek or something similar!
Well damn - I have to say that your (new) intro is impressive! Your videos are great, I think we all would like it if you made more videos more often! I'll send you a CPU from Germany if you vow to make videos more often :)
One of the applications that probably would support this is Wolfram Mathematica. Wolfram Research is listed as one of the companies that produces compatible software, and Mathematica was their flagship product. Definitely worth checking out if you ever come back to this channel!
The fact that the Weitek 3167 even existed as an alternative was a motivating factor for inclusion of an FPU in the 486DX. If it was standard, Intel figured software would be compiled for it (and it was), largely driving Weitek out of the market because the "free" FPU was plenty good enough for most purposes and very well supported because it could safely be assumed to exist. By the time the 486sx came out without the FPU, the war was already over. Intel has never liked being shown up on their own turf, and has always shown itself willing to use strong-arm tactics to avenge such losses. Just ask AMD.
Thanks for this excellent footage of a processor i never heard of before. Very interesting stuff. Impressed by the speed they achieved compared to Intel. I am always curious with what kinda video footage you show up with next. Cheers!
This is awesome! I remember reading about this chip in Byte Magazine back in the 80s or early 90s, but never been able to see one in real life. Really cool chip. Many thanks to the viewer who lent you the chip, and you for showing it running. :)
Have you ever seen a computer with a 385 cache controller? Back in the day I shortly worked on a Digital workstation with 385/386/387 chips in it and the whole machine was bizarre. Tons of extra cards for memory and such and some form of internal early SCSI for the hard drive. The original 386DX didn't have L1 cache and I assume that's what the 385 provided in an efficient fashion. Later caches were incorporated to motherboard controllers? I always wondered how much computational capacity boost it provided.
Had never even heard of that so far. Awesome. DEC was generally smart with adding enough cache to their systems (though I only had much later ones to look at)
I just found some stuff about it. It seems to have upgraded the cpu cache from 32KB to 64KB. Incredible that this worked as a discrete component back then.
Neat. I pulled a Weitek chip like this out of a Sun box about 15 years ago. I figured it was Sun-specific so I think it just ended up in a box of curiosities. Maybe I'll scrounge around and see if I can find it... and then maybe some software for it.
Thanks for the interesting video! I have the very same unopened package of a 3167-025 here but never used it in my 386 PC. There was another way to use both the 387 and the Weitek chip at the same time. There were little daughterboards available that were used to put both chips on the socket for the coprocessor at the same time. I have a Weitek 3002-0052-0 here but unfortunately one of the last owners desoldered some parts and also the pins were bent.
Great video! I Wish I had one of these Weitek coprocessors (or a Cyrix EMC87) and make a custom build of FastDoom and check if it makes the framerate even better!
Very interesting! I was always curious about the Weitek co-processor back in the day. Now I know more about it. Thanx! PS: I just found out about your channel a few weeks ago and binge watched all your videos! 😄
Yet another chip that I've always opted to only observe from a distance, indeed due to the lack of software that uses them. Very nice to see one running and I like your pronunciation of "Weitek" too, "Vytek". I don't suppose your testing with the RapidCAD was too far off, given plenty of early 486 boards had a Weitek socket (in that case, for a 4167, of course) so you kinda ended up with a scaled-down version of that. Mildly curious as to how the 3167 and 4167 compare.
My first PC was a 8088 10mhz in turbo mode with 8087. I didn't know how lucky I was until a friend with a 286 copied some games for me. When he saw them play on my 8088 he was blown away. He had more ram but the games he gave me benefitted from the 8087 math coprocessor. I think Prince of Persia was one of the games.
Came accros your youtube channel and found this very interesting video about the Witek CoProseccor - Fun fact. Olivetti actually made a computer that holds both the Intel x387 socket as well as a socket for the Weitek - Olivetti M380-40 - Intel 80386DX - On my channel I show this computer. Keep up your very interesting content. You have now a new subscriber.
The only experience I had with one was a fella who used one for code compiling. We tried to do a decent comparison to my fasmath which i ran at 33mhz to his 25, but we were unable to find software that could give a decent representation. Anecdote at the time was weitek optimization was all over the place. On some software packages it was not much better or even worse than an intel x87 and on others it may be 3x. And what small bits we found to compare were either unable to work properly on my Fasmath or didn't show his weitek in a proper light. Some code compilers of the day hated my Fasmath and would often crash. I also remember reading at some point of a patch for the windows 3.1 version of photshop for using a weitek but good luck finding it. And photoshop has very mixed code anyhow so its gains over an x87 may not be all that great... I think I am willing to concede that the benchmarks provided from weitek are probably not cherry picked to a big degree, and pretty honest. And it would be fair to say that under fairly normal conditions the weitek at about 2x of a 387, give or take. What few usable results me and the other guy managed to get seemed to show this. Our particular case showed about %50 advantage to him over my particular setup.. I also had more L2 Cache than he did if i recall. So very impressive achievement to say the least. But from my observation the Weitek was pretty much DOA. As it had no meaninful software support in its first year and its price was hard to justify over a 387 when you include the loss of compatibility and inability to run both at the same time on all but really specialized workstations. Most design artists I knew used more than one application which required hardware floating point. And within very short order the 486DX came out which pretty much sealed its fate. It never stood a chance. Its only real success was with some institutions which latched onto it from day 1 and developed their own extremely specialized and tight software stacks around it while still needing x86 compatability. And when they had the will and talent to optimize for it, you got a 3-5x increase in performance over what was possible on a 387. And such a thing would have been the best bang for the buck for at least a couple of years. It may have even been possible to also build workstations that could make use of several of them in concurrency. Though I do not know of any that did.. By that point the cost would have probably been prohibitive and you would have been within reach of a MIPS system with similar performance.
Good history. I have a Compaq 386/20 and installed a daughter board with a 387 and Weitek on it. I seem to recall around a 5x speed improvement for a chain of multiply/divide instructions.
@@billr3053 The weiteks performance seemed to vary quite a lot depending on the kind of code being ran and how optimized it was. More so than the x87. That was either a hallmark of its potential, or a signal of how bad early x87coprocessors were in peak FPU throughput. I didn't know they could come on a card? I wonder if it had any sort of local memory or fast cache to help improve its performance, as going through the ISA bus would have hobbled it quite a bit. Or wait, those early compaqs had a type of propitiatory local processor bus that used cards didn't they? Including for ram cards and what not? Some of the super early 386s also used the 287 or the 287xl due to lack of 387 being available at launch of the 386, and they were a particularly poor performer.
@@wishusknight3009 Yes, the daughterboard came with the Weitek. You take out the 80387 from the motherboard and place it on the daughterboard alongside the Weitek. Then the daughterboard plugged into the 80387 socket. There was no additional memory. The instructions to the Weitek consisted of weird looking code - simple moves to and from memory at certain locations - which made no sense just looking at the code. Its address bits being the op code. So in effect making the Weitek a memory mapped device. Cache wouldn't have helped since addressing in this way might as well be random. The toggling of address bits is what mattered and not data in memory (which wasn't present at those crazy 'locations' anyhow).
@@billr3053 Very interesting.. So it was some sort of socket extending sub strait you could use on any 68 pin socket? or did it still need to be a weitek socket on the host motherboard? I actually did no know this.
@@wishusknight3009 I don't think there was anything special on the motherboard to accommodate this. Perhaps the motherboard was a Compaq add-on as opposed to a Weitek standard. I can't find any pics online but there are some mentions of the motherboard. "There are little daughter boards available though that fit into the EMC socket and provide two sockets, an EMC and a standard coprocessor socket." See this doc: www.bitsavers.org/components/weitek/dataSheets/WTL-3167_80386_Floating-Point_Coprocessor_Sep88.pdf
I have a Magitronic 486-based system which actually happens to have the required 142-pin socket to support the Weitek 4167. While the original CPU was upgraded to an Am5x86-P75 with an interposer, im debating whether to fill the 4167 slot with the Weitek, just for completionist’s sake. If you want the motherboard model, it’s MAGITRONIC A-B435.
I am 95% sure that Autodesk 3D Studio 2 supports weitek and that alluring install option was what triggered my curiosity back then. Plenty of opportunity for nice real-world render examples if you wanna revisit the topic one day. Also, I believe the weitek only supported single-precision floating point which may have hindered its popularity as it often is not enough for engineering/science (but fine for ray tracing, wire frame drawings, etc... at least back then). This could possibly also explain some of the speed increase as the weitek would then only need to process a 24 bit mantissa (32-bit float), whereas the x87 used full 64 bit mantissa internally (80-bit float) for all operations. Thank you for the video.
would been nice to see what comes in the box, but yeah such a rare thing such be left alone, if you could run the RapidCAD cpu and Weitek copro at 33mhz then you'd probably have the best most expensive 386 back in its day, where the limited software support for the weitek fails u can still fall back on the 387 in the rapidCAD cpu.
Is that cool RapidCAD compatible with common boards of the same socket? That's a awesome chip. I've seen that chip on your channel before, but I didnt know that's what it had inside.
Remember that the 387 adds support for a lot of stuff which won't run on stock 386 hardware, like trackers, many unix oses, music software and some more recent dos games
How about that...I may at one time, when I was young, messed with a 386 machine set up like this because I do believe I have booted a PC computer that showed the "Present" for the 387 FPU and the name Weitek pop up in the boot up.......Um.... sweet! I love this channel :)
There were daughterboards that made it possible to attach a 387 as well as the Weitek chip to the socket at the same time. I got one when i bought my Weitek from eBay, but unfortunately it has bend and missing legs and some components, perhaps SMD caps are missing.
I'm not entirely sure you can trust the Weitek software's results when comparing to a i387. When I worked at Intel, they had a benchmark program which clearly showed that Intel's CPUs were faster than AMDs, this was not the case in reality.
Great video! So many coincidences in one video :D I'm currently trying to reverse engineer a GAL on one ASUS mainboard, which is responsible for Weitek 3167 binding, and some seconds later you show a mainboard with Symphony SL82C362 Chipset. Just three days ago I accidentally destroyed my favorite 386 minboard (don't ask, I just made a very stupid thing) with exactly this chipset and I'm currently desperately trying to find another broken mainboard with this chipset to transplant it to my broken mainboard and get it back to life. Actually, from my experience, it is the fastest 386 chipset I know and it is super stable...... Btw. check your mail spam folder ;)
@@5roundsrapid263 Yeah, mistakes happen, but stupid mistakes hurt even more if you repair hardware since years. At the moment, I new, that what I'm doing is a bad idea and I should be careful, but I ignored my inner voice and just moved on to fry the mainboard a minute later. I wouldn't be so sad about it, if it would be a new board, I would be able just to buy another one, but with parts, which are 30 years old it's not that easy.... :(
So thats what that bios option was for. if it was compatble with 8087 instructions i'm guessing it would be more commonlly available but performance probably wouldn't be as good
I think when it comes to FPU support a lot of people mention autocad all the time. But I think wolfram mathematica should have also been quite common and is often overlooked.
Weitek made great stuff for the workstation. They made a sun upgrade called SPARC POWER μP that made the graphics engine 2x faster. They also did powervr which is what a lot of arcade hardware used. The 3do had the powervr chip. It is probably why it was so expensive. Then 3dfx and SGi screwed it up for them.
I dug up a strange card from AST that has both a socket for the 387 as well as the enhanced socket from the looks of it. Part number is 202347, you might get somewhere with that.
I wish I had the money to pursue this type of thing, I adore all the old X86 stuff but man the prices for things now are nuts and I live in a nation with the most idiotic laws that prevent people from pulling stuff out of their local dump before it all goes to landfill, Skips, full of archaic machines, Korg/Roland Keyboards, it has all been within 12 feet of me when dropping off my own rubbish, but they wont let you touch so much as a screw of it.
Nice 387 replacement. But too bad they did not release the source code for their test programs. You never know what optimization options they used on the 387 code to make the 3167 look better. "Trust but verify".
Like others, I'd like to know the cost difference not only between the Intel 80387 and the Weitek but also the Motorola 68881/2 and the Weitek. Of course, you'd probably have to get some Sun hardware to do a performance comparison on 680x0 hardware between the 68881/2 and the Weitek since FPU support in general on the Amiga, Apple Mac, and the Atari ST/TT/Falcon was rather lacking. Maybe there was a NeXT option back when they were still in the 680x0 hardware game.
13:00 We can see here how ZIF CPU Sockets was a huge improvment for frequent changing processors, no force, no tools, just put and click and also in reverse of course. ;)
The reason you don't need the RapidCAD-2 is because only software that looks for the 287 FERR signal would be affected by it not being there. That's quite literally all it does.
Both BIOS and Sysinfo report RapidCAD-1 as 80386 but wiki says it is 80486DX. Is there a way to check this CPU for particular 486 support (extended opcodes)?
It MAY be possible to emulate a 387 on a Weitek abacus. When the 386 encounters an FPU instruction without a 387 being installed, it fires an interrupt. If you handle this interrupt with a TSR, you can issue the equivalent Weitek instruction.
Brilliant idea! yes, that should be possible, but youd need someone with enough knowledge of the weitek instruction set to write the code and the proper development tools to generate the code. Anyway, if you could find somebody to write the code, it should work. If I'm not mistaken, I believe there were compilers that would generate code that would detect if an X87 coprocessor was present of not and if not would just emulate the x87 instructions on the main processor albeit a lot slower. So, if that was possible then surely this would be as well.
I wonder if I can get a Zen 3 R9 and at to it a Weitek Abacus. I also wonder if designers of the 387 DX thought it was just as important to save memory even when they add floating point power.
Man, as a younger cat who first got into computers around 2010, seeing all of these old school CPUs with ZERO cooling solutions literally being removed with a miniature crowbar is otherworldly! haha i love this channel, so rich in PC history!
Wow.. that trident 512k splash brings back so many good memories!
Thank you to your subscriber who lent you the chip. Now we can have our cake and eat it. By the way having the "487" and the weitek at the same time was a nice fix. And one of the rarest setups of all time.
You are answering all the questions I had back in the day
^^^^^^ THIS EXACTLY
Same with me but with a delay of 25+ years ...
I never heard of anything similar back in these days. Thanks, Internet.
nostalgic..ran autocad 9 & 12..with this os.😁😁😁
The absolute undisputed #1 channel for all CPU knowledge!!!
Thank you very much! I try my best to create interesting content as good as possible.
I always wondered what the Weitek option was in the BIOS of old 386 machines! Thank you for explaining it. I was also curious how the performance of those benchmarks are on 486 and Pentium machines. Had intel already passed Weitek with their newer chips?
Yeah, with clock double technology on 486 machines Weitek was out of the game and the Intel FPU was faster. I am still struggling to get a 4167 Weitek to proof that.
I remember the days when these processors were for sale. I think they had some double or triple the price of a 387. And the 387 was not cheap either. Both chips were absolutely necessary for CAD, as it took hours to turn a complicated wire model of a cathedral 180 degrees without a coprocessor.
New intro is so.... retro :D Fits the topic of the channel :D
The March 1988 edition of Byte magazine has as the main subject and discusses the Weitek and includes the C source for the beachball benchmark.
Awesome video. As a kid growing up with Weitek in the BIOS this explains a lot. Still, a Am386 DX40 with Cyrix FasMath 387 would put that Weitek to shame. History is fascinating.
As a kid, I was always wondering, what that "Weitek" option was for. Now i finally seen this chip in action. Thanks a lot! :)
One of the few times a Cyrix was actually faster than its competitors LOL
I'm curious if the Phong demo will speed up with the rapidcad. (such as the weitek runs faster because of the better cpu)
Love the new video intro. Intriguing set of chips and quite attractive. I would love to see it working in a CAD programme. Must have been the hot-shot chip to have in those days. 👍😊
Great video, I was not even aware of this chip.
Nice combo with the RapidCAD too.
I want one of these PGA removal tools.
I want that tool too!!!
I wonder if it could be copied and 3-d printed? I'd love to have one.
This makes you wonder if there would’ve been any software available to take advantage of two FPU’s? Because as demonstrated, this setup is capable of parallel processing - in fact, you could say it is a 387 times three.
Probably the only program in existence to use both at the same time was the mandel benchmark that came with the Weitek.....
Somebody could probably hack SimCity to have a Weitek version, just as a proof of concept.
ruclips.net/video/nytpyJhDOHo/видео.html
Quake demo on a Cyrix Cx487s is about 1fps, using both FPU's in this case would give outstanding 3 fps ! Well ... I definitely wouldn't play it, but I'd love to see it as a prove of concept. ;P
Wow... I never knew you could run both co-processors.
A great video as usual 😎
You could have tested the 67fonge and the 87fonge in the Rapidcad+Weitek configuration in one go, without swapping chips.
Also it is always a bit suspicious when the test and comparison software is from the company who wants his product to be the superior, too bad you could not find a 3rd party test program.
" THE SPEED OF A $40,000 WORKSTATION, FROM A PC."
Fascinating video. I remember seeing the Weitek advertised but never realised it was a completely different type of FPU.
Ahh the FPU. I wanted one for my 486SX and didnt even know why. The first time that I actually saw the need for an FPU was when trying run Quake.
Very kind of Stefan to lend you the chip.
Someone else mentioned the oscillator already :-)
Your voice is so calming that it helps me chill out when I’m stressed. Thank you for making such amazing content
I love this channel!
You got me with the box opening though, I thought maybe you were about to do an LGR style opening and cause all sorts of collector rage :)
I'd have loved to see that!
This video is really interesting, congratulations. It reminds me of how great it was to work with ms-dos and what a computer really is
Thanks for the memories. I have a Compaq 386/20 and installed a daughterboard with a 387 and Weitek on it. I seem to recall around a 5x speed improvement for a chain of multiply/divide instructions. I did not have any compatible software for it so I wrote some test programs in assembly. One was a Mandelbrot generator. It was about 4 to 5x the speed of the 387. I didn't go any further with practical uses. Just a hobbyist around 1986.
Maybe you'll have the time to write a sample test for double FP precision for comparison with 387. ;-)
I'd never get one of these boards but it did not came with the 3167 package. It was more an add-on to support both FP accelerators at once for special purpose. Did you have the company name by hand?
@@opensparcbox I think I did those kinds of tests back in the day. The claimed 4x to 5x speedup was confirmed. I was able to achieve 90% of the theoretical throughput in my little Mandelbrot generator. The 387 did 80-bit double-extended precision internally. The Weitek only did add, subtract, multiply, divide, and square root. Not transcendental functions such as SIN, COS, TAN, ATAN. So to get full floating-point support I'd have to write all those libraries and that wasn't something I was interested in.
@@billr3053 Thanks. I appreciate your answer.
The 387 do 80-bit precision internally for all computations. So that should be near to equal, independent on the data type used (float, double). The 3167 on the other side using 31 float regs, each 32Bit wide which could be paired up to 15 double regs 64Bit wide (edit: one always has value zero, so only 15 double prec reg pairs) but then with computational performance degration, by just using supported instructions in hardware. So I'm wondering: a 4x to 5x speed improvement seems to be realistic for float type based computations only. Do you still have the assembly sources from this Mandelbrot generator? Would be nice take a look. :-)
@@opensparcbox Yes the 'single precision' was faster than 'double' on Weitek. As for the sources, hmmm. I can fire up the ol' Compaq and hopefully find those including the mapping and macro stuff I did to make programming easier. I remember optimizing and the way multiplication was done with so-called complex numbers was very symmetrical with crossing lines to the various registers. All partial results were kept on-chip.
Awesome video! The Weitec FPU at only 25Mhz really surprised me on how much faster it really is compared to that Intel 25 MHz as well. I can only imagine how much faster that FPU chip would be if it could have been used at the 33 megahertz or even a 40 megahertz version was available at the same time Intel or AMD making a chip like Intel's rapid CAD that could have been at 40 megahertz instead of just the AMD 386 DX 40Mhz alone since that chip does not have a FPU in it. Even the Cyrix 386 to 486 66Mhz upgrade and a Weitec FPU running at 66mhz (would need a 133 MHz Crystal)would have been awesome. Sad that Weitec chip did not have better software support other then mostly CAD like software. That really could have really pushed the 386 even more(the ultimate 386) during the 486 days, but sadly it was not ment to be. Great video and thank you.
I noticed that SI identified the RapidCAD as a 386SX, though... so it's limited to 16MB of memory on a 16-bit path? Sucky. Then again, even CAD software back in the day rarely used more than 16MB of memory, unless you were trying to do a very complex design... but I would think that the memory bandwidth cut might be an issue with general performance. Hardly a "486DX"... more like a 386SX/387 combo on a single socket.
As far as I know, math coprocessors in general had very little support back in the day, and the Weitek especially so. I can't think of much other than CAD, raytracing, or fractal generation, that supported a math coprocessor. Would have been nice to see some ACTUAL CAD running, or even do a run with POV-Ray (the raytracing software that you use if you're "pauvre" because it's free ;-) ). That could also have been run on a newer processor like a Pentium for reference (which I'm sure would blow it away, but just for perspective).
I don't think he tried hard enough to "make some testings," and for that I give a thumbs down.
I always have to laugh when I see "Keyboard Error! Press F1" xD
Same since PC-AT time, when one KB was switched between test bench and my main PC... no KB connected and please press F1. Good one was so AT KB may be plugged hot without powering OFF PC. PS2 Mouse always required restart.
I'm amazed that you're able to find all these rare chips seemingly with ease! Another great video!
It all depends on your budget. For the right money you can find anything.
This video explains something that puzzled me many years ago! I was at a computer fair in the early 90's and I saw one of the sellers had some random motherboards one of which looked like it had 2 processor sockets on ir... I asked the seller about it, and he said "no, its for the maths chip" which at the time I thought was strange because to my knowledge at the time coprocessor sockets were smaller than the processor socket. So probably this board was one for the Weitek or something similar!
Well damn - I have to say that your (new) intro is impressive! Your videos are great, I think we all would like it if you made more videos more often! I'll send you a CPU from Germany if you vow to make videos more often :)
thank you very much! okay. I will do my best. ☺️
Great video again, I Weited for this!
One of the applications that probably would support this is Wolfram Mathematica. Wolfram Research is listed as one of the companies that produces compatible software, and Mathematica was their flagship product. Definitely worth checking out if you ever come back to this channel!
I was waiting for that comparison since 30 years! Ty!
Great video. You answered questions which I have since decades in my head. lol
The fact that the Weitek 3167 even existed as an alternative was a motivating factor for inclusion of an FPU in the 486DX. If it was standard, Intel figured software would be compiled for it (and it was), largely driving Weitek out of the market because the "free" FPU was plenty good enough for most purposes and very well supported because it could safely be assumed to exist. By the time the 486sx came out without the FPU, the war was already over.
Intel has never liked being shown up on their own turf, and has always shown itself willing to use strong-arm tactics to avenge such losses. Just ask AMD.
Thanks for this excellent footage of a processor i never heard of before. Very interesting stuff. Impressed by the speed they achieved compared to Intel. I am always curious with what kinda video footage you show up with next. Cheers!
thanks. yeah, I try my best to show you all interesting stuff which is not so common. 😉
The design - on the box and even the chip itself - is a delight.
This is awesome!
I remember reading about this chip in Byte Magazine back in the 80s or early 90s, but never been able to see one in real life. Really cool chip.
Many thanks to the viewer who lent you the chip, and you for showing it running. :)
I remember these. We had them in the high end Sun workstations. Thanks for the memories.
700+ views, no thumbs down! Quality indeed
Have you ever seen a computer with a 385 cache controller? Back in the day I shortly worked on a Digital workstation with 385/386/387 chips in it and the whole machine was bizarre. Tons of extra cards for memory and such and some form of internal early SCSI for the hard drive. The original 386DX didn't have L1 cache and I assume that's what the 385 provided in an efficient fashion. Later caches were incorporated to motherboard controllers? I always wondered how much computational capacity boost it provided.
Had never even heard of that so far. Awesome. DEC was generally smart with adding enough cache to their systems (though I only had much later ones to look at)
I just found some stuff about it. It seems to have upgraded the cpu cache from 32KB to 64KB.
Incredible that this worked as a discrete component back then.
I recall these Weiteks were very expensive, what would the price have been back in the day? And how much cheaper was a 387?
The placement of the 8th memory slot looks funny. :)
yeah, i was also wondering. one socket misaligned. looks somehow unsatisfying 😅
@@CPUGalaxy It triggers my OCD as well..
Super interesting video and well done! The "Weitek" was mentioned in an game "StarBlade", if i remenber correctly.
That into feels like I'm watching a period correct documentary on these kinds of chips, very fitting.
Neat. I pulled a Weitek chip like this out of a Sun box about 15 years ago. I figured it was Sun-specific so I think it just ended up in a box of curiosities. Maybe I'll scrounge around and see if I can find it... and then maybe some software for it.
Thanks for the interesting video! I have the very same unopened package of a 3167-025 here but never used it in my 386 PC. There was another way to use both the 387 and the Weitek chip at the same time. There were little daughterboards available that were used to put both chips on the socket for the coprocessor at the same time. I have a Weitek 3002-0052-0 here but unfortunately one of the last owners desoldered some parts and also the pins were bent.
Very nice VLOG !! I will not be the only one when I write that we want more !!! :) and I am glad that I am not alone in being crazy about old CPUs
thanks. yeah, you are definitely not alone. 😉.
OMG so early! But I really love your new intro!
Again, very well done! ☆☆☆☆☆
Thanks for sharing. :-)
With best wishes, St.
Great video! I Wish I had one of these Weitek coprocessors (or a Cyrix EMC87) and make a custom build of FastDoom and check if it makes the framerate even better!
Wow, that ball looks exactly like in the AIBB Benchmark for Amiga!
Very interesting! I was always curious about the Weitek co-processor back in the day. Now I know more about it. Thanx!
PS: I just found out about your channel a few weeks ago and binge watched all your videos! 😄
Thank you 😊
Yet another chip that I've always opted to only observe from a distance, indeed due to the lack of software that uses them. Very nice to see one running and I like your pronunciation of "Weitek" too, "Vytek". I don't suppose your testing with the RapidCAD was too far off, given plenty of early 486 boards had a Weitek socket (in that case, for a 4167, of course) so you kinda ended up with a scaled-down version of that. Mildly curious as to how the 3167 and 4167 compare.
My first PC was a 8088 10mhz in turbo mode with 8087. I didn't know how lucky I was until a friend with a 286 copied some games for me. When he saw them play on my 8088 he was blown away. He had more ram but the games he gave me benefitted from the 8087 math coprocessor. I think Prince of Persia was one of the games.
You should also try the 87fonge/etc test to directly compare the 387 with the RapidCAD.
yeah, true. but I just got during the video the idea with the rapid cad. 😬
@@CPUGalaxy keep them for even more tests. Not in a hurry. ;-)
Came accros your youtube channel and found this very interesting video about the Witek CoProseccor - Fun fact. Olivetti actually made a computer that holds both the Intel x387 socket as well as a socket for the Weitek - Olivetti M380-40 - Intel 80386DX - On my channel I show this computer. Keep up your very interesting content. You have now a new subscriber.
The only experience I had with one was a fella who used one for code compiling. We tried to do a decent comparison to my fasmath which i ran at 33mhz to his 25, but we were unable to find software that could give a decent representation. Anecdote at the time was weitek optimization was all over the place. On some software packages it was not much better or even worse than an intel x87 and on others it may be 3x. And what small bits we found to compare were either unable to work properly on my Fasmath or didn't show his weitek in a proper light. Some code compilers of the day hated my Fasmath and would often crash. I also remember reading at some point of a patch for the windows 3.1 version of photshop for using a weitek but good luck finding it. And photoshop has very mixed code anyhow so its gains over an x87 may not be all that great...
I think I am willing to concede that the benchmarks provided from weitek are probably not cherry picked to a big degree, and pretty honest. And it would be fair to say that under fairly normal conditions the weitek at about 2x of a 387, give or take. What few usable results me and the other guy managed to get seemed to show this. Our particular case showed about %50 advantage to him over my particular setup.. I also had more L2 Cache than he did if i recall. So very impressive achievement to say the least.
But from my observation the Weitek was pretty much DOA. As it had no meaninful software support in its first year and its price was hard to justify over a 387 when you include the loss of compatibility and inability to run both at the same time on all but really specialized workstations. Most design artists I knew used more than one application which required hardware floating point. And within very short order the 486DX came out which pretty much sealed its fate. It never stood a chance.
Its only real success was with some institutions which latched onto it from day 1 and developed their own extremely specialized and tight software stacks around it while still needing x86 compatability. And when they had the will and talent to optimize for it, you got a 3-5x increase in performance over what was possible on a 387. And such a thing would have been the best bang for the buck for at least a couple of years. It may have even been possible to also build workstations that could make use of several of them in concurrency. Though I do not know of any that did.. By that point the cost would have probably been prohibitive and you would have been within reach of a MIPS system with similar performance.
Good history. I have a Compaq 386/20 and installed a daughter board with a 387 and Weitek on it. I seem to recall around a 5x speed improvement for a chain of multiply/divide instructions.
@@billr3053 The weiteks performance seemed to vary quite a lot depending on the kind of code being ran and how optimized it was. More so than the x87. That was either a hallmark of its potential, or a signal of how bad early x87coprocessors were in peak FPU throughput.
I didn't know they could come on a card? I wonder if it had any sort of local memory or fast cache to help improve its performance, as going through the ISA bus would have hobbled it quite a bit.
Or wait, those early compaqs had a type of propitiatory local processor bus that used cards didn't they? Including for ram cards and what not? Some of the super early 386s also used the 287 or the 287xl due to lack of 387 being available at launch of the 386, and they were a particularly poor performer.
@@wishusknight3009 Yes, the daughterboard came with the Weitek. You take out the 80387 from the motherboard and place it on the daughterboard alongside the Weitek. Then the daughterboard plugged into the 80387 socket. There was no additional memory. The instructions to the Weitek consisted of weird looking code - simple moves to and from memory at certain locations - which made no sense just looking at the code. Its address bits being the op code. So in effect making the Weitek a memory mapped device. Cache wouldn't have helped since addressing in this way might as well be random. The toggling of address bits is what mattered and not data in memory (which wasn't present at those crazy 'locations' anyhow).
@@billr3053 Very interesting.. So it was some sort of socket extending sub strait you could use on any 68 pin socket? or did it still need to be a weitek socket on the host motherboard? I actually did no know this.
@@wishusknight3009 I don't think there was anything special on the motherboard to accommodate this. Perhaps the motherboard was a Compaq add-on as opposed to a Weitek standard. I can't find any pics online but there are some mentions of the motherboard. "There are little daughter boards available though that fit into the EMC socket and provide two sockets, an EMC and a standard coprocessor socket."
See this doc: www.bitsavers.org/components/weitek/dataSheets/WTL-3167_80386_Floating-Point_Coprocessor_Sep88.pdf
Seeing the Present, Weitek 1167 screen was the equivalent of watching the T-800, Sarah, and John Connor entering Cyberdyne Systems
I have a Magitronic 486-based system which actually happens to have the required 142-pin socket to support the Weitek 4167. While the original CPU was upgraded to an Am5x86-P75 with an interposer, im debating whether to fill the 4167 slot with the Weitek, just for completionist’s sake.
If you want the motherboard model, it’s MAGITRONIC A-B435.
Du hast echt coole Videos, immer weiter so!
I am 95% sure that Autodesk 3D Studio 2 supports weitek and that alluring install option was what triggered my curiosity back then. Plenty of opportunity for nice real-world render examples if you wanna revisit the topic one day. Also, I believe the weitek only supported single-precision floating point which may have hindered its popularity as it often is not enough for engineering/science (but fine for ray tracing, wire frame drawings, etc... at least back then). This could possibly also explain some of the speed increase as the weitek would then only need to process a 24 bit mantissa (32-bit float), whereas the x87 used full 64 bit mantissa internally (80-bit float) for all operations. Thank you for the video.
Love these tech vids. 👍
would been nice to see what comes in the box, but yeah such a rare thing such be left alone, if you could run the RapidCAD cpu and Weitek copro at 33mhz then you'd probably have the best most expensive 386 back in its day, where the limited software support for the weitek fails u can still fall back on the 387 in the rapidCAD cpu.
Aha the weitek edition, like i mention that in the i387 video.
Thank you.
😉
Is that cool RapidCAD compatible with common boards of the same socket? That's a awesome chip. I've seen that chip on your channel before, but I didnt know that's what it had inside.
Remember that the 387 adds support for a lot of stuff which won't run on stock 386 hardware, like trackers, many unix oses, music software and some more recent dos games
How about that...I may at one time, when I was young, messed with a 386 machine set up like this because I do believe I have booted a PC computer that showed the "Present" for the 387 FPU and the name Weitek pop up in the boot up.......Um.... sweet! I love this channel :)
There were daughterboards that made it possible to attach a 387 as well as the Weitek chip to the socket at the same time. I got one when i bought my Weitek from eBay, but unfortunately it has bend and missing legs and some components, perhaps SMD caps are missing.
I'm not entirely sure you can trust the Weitek software's results when comparing to a i387.
When I worked at Intel, they had a benchmark program which clearly showed that Intel's CPUs were faster than AMDs, this was not the case in reality.
Great video!
So many coincidences in one video :D I'm currently trying to reverse engineer a GAL on one ASUS mainboard, which is responsible for Weitek 3167 binding, and some seconds later you show a mainboard with Symphony SL82C362 Chipset. Just three days ago I accidentally destroyed my favorite 386 minboard (don't ask, I just made a very stupid thing) with exactly this chipset and I'm currently desperately trying to find another broken mainboard with this chipset to transplant it to my broken mainboard and get it back to life. Actually, from my experience, it is the fastest 386 chipset I know and it is super stable......
Btw. check your mail spam folder ;)
We’ve all made mistakes before. I ruined a perfectly good Socket 7 motherboard back when they were new.
@@5roundsrapid263 Yeah, mistakes happen, but stupid mistakes hurt even more if you repair hardware since years. At the moment, I new, that what I'm doing is a bad idea and I should be careful, but I ignored my inner voice and just moved on to fry the mainboard a minute later. I wouldn't be so sad about it, if it would be a new board, I would be able just to buy another one, but with parts, which are 30 years old it's not that easy.... :(
That opening SoundFX though. Hnnnnggg.
I'm surprised a Weitek can beat the RapidCAD FPU. Even more surprised that a 3167 can work together with it.
So thats what that bios option was for. if it was compatble with 8087 instructions i'm guessing it would be more commonlly available but performance probably wouldn't be as good
I think when it comes to FPU support a lot of people mention autocad all the time.
But I think wolfram mathematica should have also been quite common and is often overlooked.
Weitek made great stuff for the workstation. They made a sun upgrade called SPARC POWER μP that made the graphics engine 2x faster. They also did powervr which is what a lot of arcade hardware used. The 3do had the powervr chip. It is probably why it was so expensive. Then 3dfx and SGi screwed it up for them.
i really love that new intro
I dug up a strange card from AST that has both a socket for the 387 as well as the enhanced socket from the looks of it. Part number is 202347, you might get somewhere with that.
Oh, new intro! ❤️
AutoCAD r12 for dos is the copy of AutoCAD you were most likely looking for. Far as I remember, it supports the Weitek Abacus.
Beautiful boxed example! Where did you find that?!?!?
Thank you so much for this video 👍🏻👌🏻✌🏻
I wish I had the money to pursue this type of thing, I adore all the old X86 stuff but man the prices for things now are nuts and I live in a nation with the most idiotic laws that prevent people from pulling stuff out of their local dump before it all goes to landfill, Skips, full of archaic machines, Korg/Roland Keyboards, it has all been within 12 feet of me when dropping off my own rubbish, but they wont let you touch so much as a screw of it.
Nice 387 replacement. But too bad they did not release the source code for their test programs. You never know what optimization options they used on the 387 code to make the 3167 look better. "Trust but verify".
Like others, I'd like to know the cost difference not only between the Intel 80387 and the Weitek but also the Motorola 68881/2 and the Weitek. Of course, you'd probably have to get some Sun hardware to do a performance comparison on 680x0 hardware between the 68881/2 and the Weitek since FPU support in general on the Amiga, Apple Mac, and the Atari ST/TT/Falcon was rather lacking. Maybe there was a NeXT option back when they were still in the 680x0 hardware game.
13:00 We can see here how ZIF CPU Sockets was a huge improvment for frequent changing processors, no force, no tools, just put and click and also in reverse of course. ;)
I think, if I remember good, I have somewhere a motherboard with two copro sockets , but a dip one for 287 and a square one for 387.
The reason you don't need the RapidCAD-2 is because only software that looks for the 287 FERR signal would be affected by it not being there. That's quite literally all it does.
Is your Intro new? Cause it looks good!
yes. its new. 😊
@@CPUGalaxy Very Nice!!
Thank you ☺️
Oh my god this box looks super familiar
I like your channel as well as retro hardware. But why do you show the box to us to not open it?
Both BIOS and Sysinfo report RapidCAD-1 as 80386 but wiki says it is 80486DX.
Is there a way to check this CPU for particular 486 support (extended opcodes)?
Unfortunately no. The RapidCAD only implements 386 instructions despite having a 486 design.
I had a motherboard with a socket for a Weitek. I never saw a Weitek for sale here though.
Excellent work!
My 486 motherboad has an weitek socket even it was an 486DX processor.
It MAY be possible to emulate a 387 on a Weitek abacus. When the 386 encounters an FPU instruction without a 387 being installed, it fires an interrupt. If you handle this interrupt with a TSR, you can issue the equivalent Weitek instruction.
Brilliant idea! yes, that should be possible, but youd need someone with enough knowledge of the weitek instruction set to write the code and the proper development tools to generate the code. Anyway, if you could find somebody to write the code, it should work. If I'm not mistaken, I believe there were compilers that would generate code that would detect if an X87 coprocessor was present of not and if not would just emulate the x87 instructions on the main processor albeit a lot slower. So, if that was possible then surely this would be as well.
Varitek was so powerful then today it's really amazing
Howly crud, two FPUs in same i386 system?
yeah. this is really cool 😎. I did not expect that to work.
Back in the day my dad got a 386dx 25. It had 2, 287 chips on the board, i always thought it was really weird. It was good at making fractles.
@@CPUGalaxy this would be the same with i486DX + W4167 ;-)
I wonder if I can get a Zen 3 R9 and at to it a Weitek Abacus. I also wonder if designers of the 387 DX thought it was just as important to save memory even when they add floating point power.