Nice to see someone else with that same Motherboard as well. But I only have the K6-2+ at 550Mhz. I have seen modded 570 K6-2+ CPUs due it's the same CPU with only half the L2 cache enabled but the modded one's are not cheap. Also not easy to know who's modification is good and whose is just a hack job.
still got a amd k2 533 - i could make it 550 but not anymore (15 years last time i tried that ) - I could not run new netscape ( belive it was 6 ) or ie 5? with great speeds sicne it felt like the browsers ate 200 mhz which made me buy a Duron 600 which did run those at good speeds
@@jarihaukilahti It probably was an issue with the RAM (volume) rather then CPU speed, because the K6-2/K6-3 were very powerful and in many tests even outperformed a Duron. The main difference, the main gain of the Duron, was with floating point and integer operations including 3DNow!, and memory bandwidth. So things like Quake or Flask ran a lot faster, while regular application sometimes even were a bit slower. So as long as you got yourself 128MB+ of PC-100 RAM on a board with sufficient cache to support it, a K6-2 400 upwards left little to be desired.
@@jarihaukilahti That should have been plenty, yes. Maybe the RAM timing was askew then, since you mixed up different types ? Or maybe you only had enough cache for the first 64MB, or 128MB - there were quite a few who fell into that trap. I switched from a K6-2 450 to a AMD Duron 700 and was quite happy myself - even after almost frying the poor thing with a bit too enthusiastic Peltier element experiments. Afterwards, I had to undervolt my Duron or it wouldn't even get past post. I switched to an Athlon Thunderbird 1 GHz from there - yet another rather significant boost. I wasn't too happy having lost ISA through this among other things though, so I still think back fondly to the (S)S7, even the first which was a K5-PR 133. :-)
Amazing series. You turned a decent early(ish) pentium board into a much more useful testbed. "Super socket 7" would allow for AGP and SDRAM, it would be interesting to see a comparison (clock for clock) with the updated boards. As an all in one test bed this board is now so much more useful and interesting. Great job.
SDRAM sure is nice to have, but only if the chipset is worth its salt ... my first PC had a 450MHz K6-2 on some shitty ATX board with SDRAM slots and ALI chipset that actually got LOWER scores in memory benchmarks than a pentium 200 on an old (non-super) socket 7 AT board with intel chipset and only SIMM memory. Also, I would not bet on getting much of an improvement out of AGP on socket 7 - I heard that AGP is buggy on all (or almost all) SS7 chipsets - this ALI board that I had sure was buggy as hell ...
Nice to see! Remember how 'powerless' we were back in the day, almost no way to know what parts would give you the most performance and what would be a bottleneck.
True, and in early internet days, there weren't many tech sites to share and get information from. Game benchmarks were still in the early days, and 3DMark2001 was still on the horizon. By which time I got a Duron 1300 system, soon to be followed by the Athlon XP 3200+ since I was too much of a power user.
It seems funny from today's perspective, but some of the best resources of the time were those giant computer magazines which were almost entirely ads that you had to pay $5-$10 for. Imagine that, paying for a magazine full of ads. The worst part was that even though those catalogs came out about once per month, the recommendations given in the articles would already be obsolete because the market changed by the time the product was published and put on streets.
That K6-3/500 is a really special chip! My first build was a K6-3/400 prior to the availability of the 500 variant. I rocked with it for nearly a decade before retiring it. I had a Tyan board that supported ECC memory and even ran Corsair ECC memory with a ramdrive. Good times! I would send it to you, but it has a special place in my heart and closet and still boots!
This has been a wild ride. I'm really impressed with your progress so far, and that of the community who have helped you along the way. Fingers crossed that you come up with even cooler results
Miss tinkering on these old machines. Im a bit of a tech hoarder and I still have a lot of my old setups buried in the garage or appropriated by my kids. Don’t fancy doing what you do anymore but it’s fun to watch you do your thing.
I ran a MB like yours with 75mhz bus for many years with a p200@225mhz, never had a single problem or instability, it was really like 5 or 6 years running like that with a trident 9660, i just set the jumpers to 75mhz and never put back because never had a reason to, it just worked... true the pci ran @37.5mhz but it just worked fine even with my voodoo 1 and playing games for hours non stop, at times i would ran it with bus at 83mhz/pci 41mhz but the proc at 2.5x instead of 3x giving it 210mhz and it would also work fine for hours but everything exept the voodoo card that did not work at those speeds.... along with that one i also got a asus tx97 with a p200mmx also @225mhz rock solid, this one would not go but over than 75mhz bus. A friend of mine still have his P55T2P4 with intel HX chipset, this board we tested with a k6 3 at litle over 400mhz and bus 75mhz, unfortunetly i didn´t have my old diamond stealth (s3) that would rua at 83mhz bus anymore to test the k6 even higher at that time.
Unfortunately, I don't know any 430FX mainboard, which would provide 75MHz FSB, that would require some heavier modifications. That FSB clock started to come first on later mainboards with 430VX and VIA MV2 chipset.
I've been using Jan's modified BIOS since 2001 when I read Oldie Tuning on Tom's Hardware and upgraded my AT computer with a P55T2P4 and a K6-2+ 475 (which overclocked to 500 MHz by using the 83 MHz frontside bus). Jan has always been very helpful, and has done his BIOS magic on request for myself and a number of others I know of. I've used both the 75 and 83 MHz FSB on many Socket 7 motherboards and have had no issues with the out of spec PCI bus. Last year I picked up an old AT computer for $20.00. This was a major score for me since I really wanted a compact AT case, and they have become very rare here. It came with a PCChips M520 VX chipset motherboard, 128 megabytes of EDO RAM, a K5 PR133 (100 MHz) an extremely slow hard drive, but most importantly for the poor person who was forced to use it, no L2 cache at all! This was one of the infamous PCChips motherboards that had fake cache chips, and the COAST slot was empty. On the bright side, this rendered the VX chipset's sixty four megabyte cacheable area limit moot! I used an Evergreen Spectra voltage regulator and replaced the K6-2 400 CPU that was installed in it with a K6-3+ 400 that I slightly overclocked to 450 MHz (6x75). Combined with a fast, modern hard drive, this computer flies in comparison to how it was when I first got it. Since this board has a header for a VRM module, I plan to replace the Evergreen module one with one of yours.
This is exactly the kind of thing I would do - end up testing every configuration with every CPU just to satisfy my own curiosity. Well, you also satisfied my curiosity, sir! Thank you.
My father have had this P/I-P55 with his minitower PC, I enjoyed so much to upgrade his flash BIOS with BP patcher... thank you lots for your titanic work, it is really needed for those beginners starting to understand this field of interest... thank you.
You made that board go ballistic. The wizard of upgrades! IC by Necroware (ICBN) I just do really hope you'll get your due credits for your invention of your voltage regulator. Making it open-source is a humongous credits worth of kudos. PS: Sorry for my inadequate English but I hope you understand my admiration. Kind regards
I had a Pentium 166-MMX when I was a child. I used that PC to take my first steps in the world of programming. Now I want to blow off the dust from my P1 PC and go back to the spring of my life. Thank you for the video.
I've followed this series from the beginning and it's been really fascinating. I love how much you were able to push this board. Also, I've always been curious if there was any difference in performance between MMX and non MMX pentiums in regular applications. It's cool to see that there actually was! Looking forward to see what you explore next.
Could you recompile Doom with MMX optimizations enabled? Since it has been reverse engineered and compiled for a billion platforms I'm guessing it could be done 🤔🤓 it would be REALLY interesting to see the difference in benchmarks, if any 🙏🏻🙏🏻 Thanks for another amazing video, I'm always looking forward to watching them
Hm, Doom is probably not written in a way that could benefit too much from MMX instructions. The way it renders the frames is totally not the way a SIMD instruction set would like it. It goes in columns, uses paletted 8-bit pixels, has look up operations for every single pixel and does hardly anything on groups of pixels. Moving rendering into a system memory buffer could possibly help, at least it could with slow video cards (I remember Heretic was a hell lot of faster than Doom on my very unbalanced Am5x86 + ISA Tseng ET3000 system) and MMX could even make a faster frame buffer transfer possible - probably not by much though as the 32-bit-wide PCI bus could be a bottleneck here. Hm. Audio processing and mixing could be a better candidate, I guess, but that's less of a bottleneck.
The Doom Engine source code was released by iD software, under the GPL. It was not reverse engineed. Same goes for the Engine for Quake 1, Quake 2, Quake 3, and Doom 3.
That's awesome. I remember playing with Socket 7 in the early 2010s. 200mhz pentium mmx, 512mb sdram, 512kb L2 cache module, SATA PCI controller. Got stuck trying to put PCI 3D graphics on it, and I tried upgrading to AMD K6 but the voltage was too high. A lot of fun watching you take it to the next level!
Very good video! Entertaining and well thaught. As pcb design guy i could help you with dc_dcs. Well from year 2013 I started a new hobby: extreme overclocking. I bought single-stage-phase-change-cooling and I had -42*C at 150W cpu. Very time consuming hobby, but fun. I have done "8800 zombie mod" - cut dc-dc part of pcb and mod it like it is a adjustable power module. Now I am selling my collection, but everything older than socket A is going to be kept for future projects. So there are more of us, retro fans! Ps. Overclockers used hwbot site. Ps2. The most wanted mobos were: ga-5ax and asus p5a. The goal was 2d performance, tests like superpi, pifast, wprime, cinebench, no 3d like 3dmark, aquamark or games.
Wooow 256kB cache stick /me has now a 5950X with 8MB L2 en 32MB L3 cache, we came a long way. My first pc was a 166MMX then after some pebcacks (= new mobo's) I upgraded the cpu to a K6-3/550 :) what a time
very impressed by your level of proficiency for someone who claims not to be an electronics engineer, I think that you could put a few of them to shame with your resourcefulness. This is genuinely excellent work. I've gotten the bug from you so now my girlfriend wants to kill both of us because of the number of PCB boards occupying the house!😂
I've casually watched your videos over a couple of years... perhaps over a couple drinks after work... Today you blew me away with an era I don't even dabble in (I tend towards 2003-2010 stuff, I guess). It's been awhile since I saw such detailed results, such devotion to a topic. I also more than kinda love the whole creepy vibe of the intro - I run a charity haunted house for Halloween (yes, I use retro/ewaste tech in it, how did you guess?). If I ever did a tech channel (I have so thought about it, but time, man, time) - I would do it in character, like a horror movie host... I wouldn't be able to resist! I made sure my subscription button was on, with the alerts. Thanks. This was great.
That's true, some manufacturers were so evil to use a custom pinout, but so far I only had issues with COAST modules on 486 boards. The module, which I showed in this video has the right ground and vcc pinout, so there was no chance to see the white smoke, but it still seems not to work properly though.
Excellent work! Videos like this make me wish I had held onto my old Megatrends HX83 socket 7 board and AMD K6 200Mhz CPU. I don't remember what happened to it when I replaced it with an Abit BH6 mainboard and a Celeron 300A that spent years overclocked to 464 Mhz.
Thank you! The only missing CPU in this test IMHO tillamook - as I remember its overclockable till 400MHz, so its interesting can it compete k6-iii+ on this clock.
Hey, you did some great tests. That brought me back so much memories. But I've got one question: If I remember correctly from back in the days, the RAM is cached top to down, so upper memory first. If you use more memory than cachable by the chipset, the lower memory will be uncached. The i430FX should have a cachable area of 64 MByte. Thus it should run faster with only 64 MByte RAM for anything not demanding more than 64 MByte. And that is certainly true for most DOS apps and games. Greetings Peer
I think the i430 caches the lowest 64 MB, but it was Win95 that liked to allocate top-down. (And NT or Linux would allocate all over.). Regardless, your point is correct: as long as you don't _need_ the extra RAM, sticking with
The sockets on the onboard asynchro-sram seem to be longer than the chips installed. Is there a possibility to take this up to 512KB and use it for level-3 cache with the AMD K3-500 ? Perhaps with yet another mod on the bios ? And what about the Pipeline Burst ? Any 512KB availablity there ?
Great video! Thanks! The SETMUL utility can change multiplier for AMD K6-III Plus "on the fly". On fsb = 66MHz the CPU may operate from 133MHz to 400MHz with 33MHz steps! Very awesome!
Cache makes a difference in many cases. I remember that my K6-2 overclocked @ 600 Mhz was still slower in many things (not all) than a K6-III @ 400Mhz.
COAST modules were not usually very interchangeable with each other despite using a standardized slot, it was as if the manufacturers did not want you to be able to reuse or sell them...
From what I understand it was extremely optimized on Assembler to the Pentium (non mmx) features. Who knows if optimizing for K6 and even 3dnow would improve performance much. Best wishes.
@@ajax700 It would. But 3dnow came with K6-2, not K6. And it was 1998. So ID software focused only on Q2 and later Q3, to support 3dnow!. K6 came also after Quake was released. K5 wasn't so great, in comparation with Pentium 1, so maybe they decided they have time to optimalize engine only for one of them. Maybe they underestimated AMD, and didn't expect K6 will be so good. So choosen only Pentium. And in 1997, probably decided, they will not focus on rebuild engine to K6 too, maybe time was precious, and they were already focusing on GLQuake and Quake2. Best wishes too.
@@warrax111 Amd almost never supplied more than around 30% of the cpu market, so it makes A LOT of sense to optimize for the most used CPUs. They also provided compilers, tools, detailed documentation, etc. Best wishes.
Really cool project, very interesting how far you can push this old hardware and the vrm is genius. I wonder if those coast cache modules come in different speeds and how much that would affect the system as a whole. Also would be interesting if you could find a 512k version too see if there are any benefits or downsides to that. Keep up the great work and awesome videos!
during whole video my inner voice was screaming: K6-3! put K6-3 in there! and than ... BOOM :D and second thing. I remember when I was first time inspecting my socket 7 board many years ago. I was wondering what are those various slots for. and now I know :)
Just found your channel :) Very nice. I'm just getting back into old computer hardware after a break. Have to replace some plastic 30pin SIMM slots on my 386 motherboard :( lol. But found some spare slots on a 386 backplane I have that will probably never be used again due to corrosion form the battery etc.
I liked this series of videos so much! 😊 The idea of getting a piece of hardware, a motherboard, to work in a way which it only ever had the potential to do is really appealing to me. I really love the idea of extracting the most from the options of a motherboard. I commented in the past about SMPS running at an higher frequency than the current design. Now it's probably not that interesting since the VRM seems to do its job quite well but I've done some research and there are a couple of ICs which could be used. One is the OnSemi NCP3020B (version -B for 600MHz switching frequency) in SOIC-8 package for which there is an evaluation board, NCP3020BGEVB. Schematics, bill of material, gerber files are available. Unfortunately the board is not in stock at any of the usual distributors so I'm not sure whether it has been discontinued. An alternative controller is TI LM2745/8 (adjustable from 50 kHz to 1 MHz) in TSSOP-14 package. It's slightly trickier to hand-solder and more expensive. It features a Power-Good open-drain output. The evaluation module, LM2745-19AEVAL, is in stock at Digi-Key in the UK for £75.372 including VAT.
Thank you very much! I'm trying to make my projects as inexpensive as possible. The parts for my VRM cost around 8-10€, of course developing the prototype was more expensive in average, since I had to try different parts and designs, but it is still by far not £75. Still thanks for suggestions.
@@necro_ware You're welcome. The evaluation board would be a Non-Recurring Engineering (NRE) cost for the new VRM development 😉 Unfortunately they are generally expensive and they cost quite a lot more than the cost of the parts. In the industry they help speeding up development and time is money! Even if you don't buy these evaluation boards it's good that the manufacturer of the IC makes them because they provide example of (hopefully) working schematic and layout which you can copy or use for inspiration.
@@lorenzo.c That's true! For my day time work I wouldn't go the same way, as what I do here, but here there is one absolutely free resource, which I can rely on - my time ;) Furthermore, most projects which I do are to learn something, that's why also in this case I didn't want to just copy something, but calculated everything from scratch. I really wanted to understand each click which happens in this circuit. One additional challenge which I always like to set is to be cost effective in terms of production. Of course, my engineering time is not included and that is the most expensive resource in the industry, but in the hobby the journey is the reward.
Wow , didn't expect so big differance between K6-3+ Asynch cache, and PB cache. K6-3+ has internal L2 cache, which is very fast and ticks at processors speed, so in this case 400mhz. L2 cache from motherboard become L3 cache, and it's used only sporadicaly, when CPU runs out of L2 cache. Seems Quake can take all CPU cache and still need more. Incredible.
Wow that Cos module made a day and night difference, never thought i would see sutch a big jump, i wonder kf we can recreate them nowwdays as i have a asus board that could benifit from one :) PS great job on the power module , I wouldnt mind picking one up at a later date :)
Must be over 25 years ago but I remember those older Asus boards with the media bus slot at the end of one of the pci slots, the p55t2p4 with the right revision I think rev 3 and above would overclock and run a bus speed of 83mhz just by setting jumpers on the board- and was the fastest board out there at the time - it was posted on Toms Hardware about 25-27 years ago but your running with a fixed clock divider! I remember it well as I bought the board for that reason, andran mine at 83Mhz up until the p2 came out without issues after I got it settled down, from memory I only run intel CPU's in the board, I had a sb16, a GUS Max and a diamond stealth 64 (S3 chipset) with an NE2000 compatible nic, all seemed to behave ok in it at that speed, I wrote a simple multi-config DOS menu that i used to assign irq and dma to the sound and scsi,I used a multi config to change settings by menu selection at boot time as needed , there were cards that didnt like the high bus speed like the AWE32 would sound wierd and I did have to swap out the 1542 adaptec scsi card for a PCI NCR one as the adaptec didnt like the overclock neither did the advansys or buslogic cards behave at overclock - great stuff it brings back great memories and always look forward to viewing your channel what you post thankyou
Awesome video, as always! I'm still sad that I could not repair my Asus Socket 7 motherboard with a broken CPU socket. The main reason being that I could not find a broken MB with a good socket and I would not want to destroy a working MB.
Thanks for this video and the always interesting content! I have to add something about the MMX Pentiums... they were a little more than just more cache and MMX. There was a better branch predictor, borrowed from the Pentium Pro. But that's not all! The Pentium has two pipelines, called U and V. The U pipeline could run all instructions in parallel, the V, however, could not. You carefully had to order your instructions in order to be paired on the original Pentium. This especially applied for prefixed instructions - those not available on the original 8086 - so basically several 32 bits instructions.The MMX Pentium, however, could handle more operations in the V pipeline, and hence, this, with the larger cache, and better branch predictor, have a notable impact on performance as well, as it could pair more instructions per cycle. I wasn't aware by HOW MUCH pipelined burst cache could speed up those old CPUs... have some stick lying around and surely one or two boards - need to try it ;)
I have the UMC burst pipe module in my m520 and it works without any issues. So I think that might be a bad module. I could test it in my PCchips m520 or if you have one you might want to try it there. Oh the m520 also has a VRM socket.
Very nice benchmarking and comparison! I ought to do similar with my Socket 7/Voodoo 1 test box. It's currently running a K6-266 because that's what was in it.
A pipeline burst cache... I totally forgot about that. It had a HUGE impact. I went for a Pentium 90 instead of the P100 an invested the money into PB cache instead. This was a great decision as my home-built P90 wiped the floor with the Compaq P120 that we used in the office. Too sad you didn't try the FSB upgrade. I totally understand the reason of course, but it would have been really interesting. Sure, not 100Mhz, that will most likely not work, but maybe 75Mhz. At least on later motherboards a moderate overclocking of the FSB to 75MHz was always stable. I also never had a PCI card that didn't work with 75Mhz. It would be interesting, if this works for the FX chipset too. Then you would get the AMD chip to 450MHz and see the impact of FSB increase.
The 512k COAST module gives a further performance boost. Probably a decade ago now, I tested the difference between a 256k and 512k COAST module on a P200 MMX, and there was a 5-7% performance improvement in Quake. There's supposedly a rare 1 MB COAST module, but I never seemed to be able to find one.
a k6-III was able to work on this socket 7 motherboard??! 😱 Wow...I never knew the potential this motherboard was capable of having. Fantastic series this was. My concern though is the build itself. Can it be put to the test with voodoo 2 (sli or not)? What atx power supply is recommended on socket 7 motherboard btw?
I enjoyed the video! I would like some details though, from someone. The last CPU (printed 500MHZ) was running at 400MHZ? 66.6 MHZ FSB bus by 6 multiplier CORRECT? And it worked faster than the others because of what? The internal cashe? That external PipelineBurst cashe module really improved things.
It would be cool to see more comparsions if you’re able to test with boards that can utilize higher fsb just to see what performance difference it gives
66MHz FSB bottlenecked all CPU's starting from around 166MHz, and that bottleneck became even worse after 200MHz. MMX-Pentium's bigger L1-caches helped a little, but underlying slow 66MHz FSB and 33MHz PCI-bus strangled the whole system. I ran 350MHz K6-2 as 336/112 because it was almost as fast as 400/100. But ran cooler with stock voltages. Slightly higher PCI-bus (112/3=37.3MHz) helped too. Practically all PCI-cards I used (Matrox/Voodoo1/3Com NICs) worked with 37.3/37.5MHz PCI-speeds, sadly many began to fail with 41.5MHz PCI-bus (83MHz FSB).
I remember wondering how the FSB speed division worked on motherboards, I didn't know enough to investigate it back then though. Do you know? is it handled by a crystal or an IC somehow? I mean the division of the FSB to PCI/ISA bus speed.
From what i remember say PIIX has all the clocks supplied externally and it doesn't care where they come from. But to convert between different clock domains, it performs buffering. So there's two parts to every multi clock domain system, for one clock generation, and for other signal translation which happens on whatever logic IC that has both clock domains, and signal translation for PCI to ISA is on the PIIX. But PIIX being a PCI device, of course it's only half the picture, it never has to deal with FSB. Let's see, i fished out a 430tx mainboard reference schematic. I can see there a dedicated clock generator IC which is part of the chipset, IMISC652 or IMISC671A. I also remember that newer boards usually use a similar multi clock domain generator IC, it's often third party, there's a bunch of suppliers, Cypress probably the main one. A tiny little thing, not noticed at a glance. So this one has a 14M318 crystal oscillator and generates PIIX clocks, 48M (bet that's the USB base clock on the PIIX), 60/66M, PCI, FSB and DRAM clock i think. I can see on this board also a 32K768 RTC crystal in either of the two places, could be on IO IC or on the PIIX. So that's it, no more crystals. Fundamentally an AM3 system still looked the same, with that 14M318 and some cryptic proprietary clock generator spawning everything from that. AM4 looks different, there's a 25MHz crystal at the chipset; and a 48MHz one and an RTC one at the AM4 socket, and i guess all the other frequencies just come out of either of the two chips, mostly the processor just generates the clocks internally a whole bunch of them. Fun fact: apparently fundamentally reliable, race-free clock domain transition between uncorrelated clocks isn't possible. I don't know if it's an actual fact, i haven't verified it first hand, but i've been told that it is so. But it's possible to have probability of failure astronomically low, and that's probably good enough for systems of the eras where the domains are independent. Today all interfaces are packet oriented which presents an extra opportunity, that packets are buffered input side, can be checksummed and re-sent, so you can cover up the potential unreliability of clock domain transition this way.
Funny misspelling :) Yeah I tried, but there are only tools which give a boost for higher resolutions, which I didn't present in this video. May be next time.
@@necro_ware Indeed :D Well, if you can turn it on for the A000-AFFF area you can see a healthy speed boost in Quake 320x200 or even in Doom (not that sure as it has a strange video memory access pattern). K6WCX surely did it for me. In a real SS7 system, I must admit, yet it should work just the same here.
Seems like Quake 1 is playable on all configurations. I've always wondered if K6-2+/3+ could reach a level of performance on Socket 7 to run games like Quake 3 and Unreal Tournament at at least 25 fps, with 3D acceleration of course. Seems like that would be quite a challenge, even with completely overpowered AGP cards.
The issue you'll run into trying to reach those performance levels, is that these chips really need the FSB speeds of an SS7 board to get there easily. Having a faster FSB allows for much higher bandwidth between the CPU and GPU. I have yet to try it, but my FIX VA-503+ board has I think a 124Mhz option for the Front Side Bus and a whopping 1MB of onboard cache. The K6-2+ I'm using and the ram seem pretty rock solid in this configuration. It's most just a matter of finding the time to try some games. I feel like what might actually bottleneck it is the Voodoo Banshee I have in there. I wonder if anyone would know if the CPU would bottleneck the card, or if the card would bottleneck the cpu?
@@Edvinas_channel Perhaps. It depends on the silicon lottery. The ram needs to be high quality and built for 133Mhz (Pentium III era ram), and the cpu needs to be able to handle the FSB. I would make sure to pair it with a card that is actually a little too fast for the cpu like a Voodoo 3 so as to remove bottlenecks. Even this, these games really aren't what the K6-2+ and better cpus were really great at. They let you run a ton of Windows and DOS games out of one machines while making a few compromises; performance on late 90s 3D games being one. That said, it should probably be possible to run UET and Q3 at 30 or better *most* of the time, but I wouldn't expect it above 640x480. I haven't tried it yet myself, so I don't actually know. Just telling you to mute your expectations.
I'm still sporting a K6-3+ at 600mhz in my GA-5AX board. Quake2 actually got a 3dnow! patch to give it a performance boost on these CPU's.
Nice to see someone else with that same Motherboard as well. But I only have the K6-2+ at 550Mhz. I have seen modded 570 K6-2+ CPUs due it's the same CPU with only half the L2 cache enabled but the modded one's are not cheap. Also not easy to know who's modification is good and whose is just a hack job.
still got a amd k2 533 - i could make it 550 but not anymore (15 years last time i tried that ) - I could not run new netscape ( belive it was 6 ) or ie 5? with great speeds sicne it felt like the browsers ate 200 mhz which made me buy a Duron 600 which did run those at good speeds
@@jarihaukilahti It probably was an issue with the RAM (volume) rather then CPU speed, because the K6-2/K6-3 were very powerful and in many tests even outperformed a Duron. The main difference, the main gain of the Duron, was with floating point and integer operations including 3DNow!, and memory bandwidth. So things like Quake or Flask ran a lot faster, while regular application sometimes even were a bit slower.
So as long as you got yourself 128MB+ of PC-100 RAM on a board with sufficient cache to support it, a K6-2 400 upwards left little to be desired.
@@l3lue7hunder12 I had even pc-133 memory , a game like unreal tournament like 160mb ram the best
@@jarihaukilahti That should have been plenty, yes. Maybe the RAM timing was askew then, since you mixed up different types ? Or maybe you only had enough cache for the first 64MB, or 128MB - there were quite a few who fell into that trap.
I switched from a K6-2 450 to a AMD Duron 700 and was quite happy myself - even after almost frying the poor thing with a bit too enthusiastic Peltier element experiments. Afterwards, I had to undervolt my Duron or it wouldn't even get past post. I switched to an Athlon Thunderbird 1 GHz from there - yet another rather significant boost.
I wasn't too happy having lost ISA through this among other things though, so I still think back fondly to the (S)S7, even the first which was a K5-PR 133. :-)
Hands down, without a doubt, THE best retro x86 channel out there. Outstanding work. Thank you so much!
Genuinely incredible work. The satisfaction on seeing those benchmark figures! love it!
Amazing series. You turned a decent early(ish) pentium board into a much more useful testbed. "Super socket 7" would allow for AGP and SDRAM, it would be interesting to see a comparison (clock for clock) with the updated boards. As an all in one test bed this board is now so much more useful and interesting. Great job.
SDRAM sure is nice to have, but only if the chipset is worth its salt ... my first PC had a 450MHz K6-2 on some shitty ATX board with SDRAM slots and ALI chipset that actually got LOWER scores in memory benchmarks than a pentium 200 on an old (non-super) socket 7 AT board with intel chipset and only SIMM memory.
Also, I would not bet on getting much of an improvement out of AGP on socket 7 - I heard that AGP is buggy on all (or almost all) SS7 chipsets - this ALI board that I had sure was buggy as hell ...
@@KenjiUmino Well, then use SIS chipset. Rare, but much faster in relative performance.
This was amazing to watch. Incredible how each piece of hardware added performance and seeing the end result was just crazy. Awesome job.
Nice to see! Remember how 'powerless' we were back in the day, almost no way to know what parts would give you the most performance and what would be a bottleneck.
True, and in early internet days, there weren't many tech sites to share and get information from. Game benchmarks were still in the early days, and 3DMark2001 was still on the horizon. By which time I got a Duron 1300 system, soon to be followed by the Athlon XP 3200+ since I was too much of a power user.
It seems funny from today's perspective, but some of the best resources of the time were those giant computer magazines which were almost entirely ads that you had to pay $5-$10 for. Imagine that, paying for a magazine full of ads. The worst part was that even though those catalogs came out about once per month, the recommendations given in the articles would already be obsolete because the market changed by the time the product was published and put on streets.
We can finally rest now, knowing what configuration was the utterly best for S7.
It blows me away. Genius series, genius results. Nothing more to say ❤
Danke für's Ausleihen der i200 noch Mal! Die war sehr nützlich bei diesem Experiment :)
That K6-3/500 is a really special chip! My first build was a K6-3/400 prior to the availability of the 500 variant. I rocked with it for nearly a decade before retiring it. I had a Tyan board that supported ECC memory and even ran Corsair ECC memory with a ramdrive. Good times! I would send it to you, but it has a special place in my heart and closet and still boots!
This has been a wild ride. I'm really impressed with your progress so far, and that of the community who have helped you along the way. Fingers crossed that you come up with even cooler results
Miss tinkering on these old machines. Im a bit of a tech hoarder and I still have a lot of my old setups buried in the garage or appropriated by my kids. Don’t fancy doing what you do anymore but it’s fun to watch you do your thing.
Great conclusion to a fascinating series.
always great to see another video from you
this board has been absolutely maxxed out and it's good to see
I ran a MB like yours with 75mhz bus for many years with a p200@225mhz, never had a single problem or instability, it was really like 5 or 6 years running like that with a trident 9660, i just set the jumpers to 75mhz and never put back because never had a reason to, it just worked... true the pci ran @37.5mhz but it just worked fine even with my voodoo 1 and playing games for hours non stop, at times i would ran it with bus at 83mhz/pci 41mhz but the proc at 2.5x instead of 3x giving it 210mhz and it would also work fine for hours but everything exept the voodoo card that did not work at those speeds.... along with that one i also got a asus tx97 with a p200mmx also @225mhz rock solid, this one would not go but over than 75mhz bus. A friend of mine still have his P55T2P4 with intel HX chipset, this board we tested with a k6 3 at litle over 400mhz and bus 75mhz, unfortunetly i didn´t have my old diamond stealth (s3) that would rua at 83mhz bus anymore to test the k6 even higher at that time.
Unfortunately, I don't know any 430FX mainboard, which would provide 75MHz FSB, that would require some heavier modifications. That FSB clock started to come first on later mainboards with 430VX and VIA MV2 chipset.
DUDE that is awesome! i found my way here as i have been pondering how to get my pb multimedia s606 to work with a k6
I've been using Jan's modified BIOS since 2001 when I read Oldie Tuning on Tom's Hardware and upgraded my AT computer with a P55T2P4 and a K6-2+ 475 (which overclocked to 500 MHz by using the 83 MHz frontside bus). Jan has always been very helpful, and has done his BIOS magic on request for myself and a number of others I know of.
I've used both the 75 and 83 MHz FSB on many Socket 7 motherboards and have had no issues with the out of spec PCI bus.
Last year I picked up an old AT computer for $20.00. This was a major score for me since I really wanted a compact AT case, and they have become very rare here. It came with a PCChips M520 VX chipset motherboard, 128 megabytes of EDO RAM, a K5 PR133 (100 MHz) an extremely slow hard drive, but most importantly for the poor person who was forced to use it, no L2 cache at all! This was one of the infamous PCChips motherboards that had fake cache chips, and the COAST slot was empty. On the bright side, this rendered the VX chipset's sixty four megabyte cacheable area limit moot! I used an Evergreen Spectra voltage regulator and replaced the K6-2 400 CPU that was installed in it with a K6-3+ 400 that I slightly overclocked to 450 MHz (6x75). Combined with a fast, modern hard drive, this computer flies in comparison to how it was when I first got it. Since this board has a header for a VRM module, I plan to replace the Evergreen module one with one of yours.
That Никс label on P1 sure brings back memories. My dad's friend used to run a local Никс shop branch.
I didn't knew this was a shop. I got this CPU donated from a viewer from Slovakia.
@@necro_ware you can read a bit about the company in the russian Wiki - ru.wikipedia.org/wiki/Никс_(компания).
K6-3 +'s are so pricey, what a find!
Absolutely, just imagine my face, as I saw it in the scrap.
This is exactly the kind of thing I would do - end up testing every configuration with every CPU just to satisfy my own curiosity. Well, you also satisfied my curiosity, sir! Thank you.
You bought 286 pc from me many years ago lol
My father have had this P/I-P55 with his minitower PC, I enjoyed so much to upgrade his flash BIOS with BP patcher... thank you lots for your titanic work, it is really needed for those beginners starting to understand this field of interest... thank you.
You made that board go ballistic. The wizard of upgrades! IC by Necroware (ICBN)
I just do really hope you'll get your due credits for your invention of your voltage regulator. Making it open-source is a humongous credits worth of kudos.
PS: Sorry for my inadequate English but I hope you understand my admiration.
Kind regards
You really have to love that platform, I have a P55T2P4 Rev 3.1 with a K6-3+ @ 400 MHz + Voodoo 1 6MB running and it is a blast :)
Blast from the past !
That Pentium 200 mmx was my childhood pc's cpu for a long time, with a voodoo 3 2000 ❤
I was in my late teens and had a K6-2 350 with the same Voodoo 3 2000, good times with Unreal tournament and Screamer 2❤
Great work on pulling all these results together!
Love this series. The days of exciting and meaningful upgrades. 👌
I had a Pentium 166-MMX when I was a child. I used that PC to take my first steps in the world of programming. Now I want to blow off the dust from my P1 PC and go back to the spring of my life. Thank you for the video.
I've followed this series from the beginning and it's been really fascinating. I love how much you were able to push this board. Also, I've always been curious if there was any difference in performance between MMX and non MMX pentiums in regular applications. It's cool to see that there actually was! Looking forward to see what you explore next.
you are the gentleman of retro-computing!
Amazing series. Thank you for sharing your journey with us!
Excellent benchmarks
I truly miss those days of pc gaming. It was magical
Could you recompile Doom with MMX optimizations enabled? Since it has been reverse engineered and compiled for a billion platforms I'm guessing it could be done 🤔🤓 it would be REALLY interesting to see the difference in benchmarks, if any 🙏🏻🙏🏻
Thanks for another amazing video, I'm always looking forward to watching them
Hm, Doom is probably not written in a way that could benefit too much from MMX instructions. The way it renders the frames is totally not the way a SIMD instruction set would like it. It goes in columns, uses paletted 8-bit pixels, has look up operations for every single pixel and does hardly anything on groups of pixels. Moving rendering into a system memory buffer could possibly help, at least it could with slow video cards (I remember Heretic was a hell lot of faster than Doom on my very unbalanced Am5x86 + ISA Tseng ET3000 system) and MMX could even make a faster frame buffer transfer possible - probably not by much though as the 32-bit-wide PCI bus could be a bottleneck here. Hm.
Audio processing and mixing could be a better candidate, I guess, but that's less of a bottleneck.
The Doom Engine source code was released by iD software, under the GPL.
It was not reverse engineed.
Same goes for the Engine for Quake 1, Quake 2, Quake 3, and Doom 3.
Doom never supported MMX.
@@Jeroensgambling
There are source ports that support all sorts of fancy things.
It used lot of tricks, like lookup tables to actually not do floating point operations, which were expensive. MMX is about floating, so...
Another high quality video
That's awesome. I remember playing with Socket 7 in the early 2010s. 200mhz pentium mmx, 512mb sdram, 512kb L2 cache module, SATA PCI controller. Got stuck trying to put PCI 3D graphics on it, and I tried upgrading to AMD K6 but the voltage was too high. A lot of fun watching you take it to the next level!
Very good video! Entertaining and well thaught. As pcb design guy i could help you with dc_dcs. Well from year 2013 I started a new hobby: extreme overclocking. I bought single-stage-phase-change-cooling and I had -42*C at 150W cpu. Very time consuming hobby, but fun. I have done "8800 zombie mod" - cut dc-dc part of pcb and mod it like it is a adjustable power module. Now I am selling my collection, but everything older than socket A is going to be kept for future projects. So there are more of us, retro fans!
Ps. Overclockers used hwbot site.
Ps2. The most wanted mobos were: ga-5ax and asus p5a. The goal was 2d performance, tests like superpi, pifast, wprime, cinebench, no 3d like 3dmark, aquamark or games.
Wooow 256kB cache stick /me has now a 5950X with 8MB L2 en 32MB L3 cache, we came a long way. My first pc was a 166MMX then after some pebcacks (= new mobo's) I upgraded the cpu to a K6-3/550 :) what a time
This series is my favorite on all of TH-cam!
this is a great video, thanks for all your work and time.. cheers.
thats one hell of an upgrade. very nice
very impressed by your level of proficiency for someone who claims not to be an electronics engineer, I think that you could put a few of them to shame with your resourcefulness. This is genuinely excellent work. I've gotten the bug from you so now my girlfriend wants to kill both of us because of the number of PCB boards occupying the house!😂
Quake rocking 66.6fps seems really appropriate 🤘
I've casually watched your videos over a couple of years... perhaps over a couple drinks after work...
Today you blew me away with an era I don't even dabble in (I tend towards 2003-2010 stuff, I guess). It's been awhile since I saw such detailed results, such devotion to a topic.
I also more than kinda love the whole creepy vibe of the intro - I run a charity haunted house for Halloween (yes, I use retro/ewaste tech in it, how did you guess?). If I ever did a tech channel (I have so thought about it, but time, man, time) - I would do it in character, like a horror movie host... I wouldn't be able to resist!
I made sure my subscription button was on, with the alerts.
Thanks. This was great.
Thank you! Yeah, I guess my videos are even more enjoyable with a drink ;)
Absolutely mindblowing upgrade. Nothing to add on top of it!
I love when the hardware is hacked so neatly (not like bodges everywhere for example) to make it do things beyond what was intended
Awesome stuff. One day I will have to dive into the science of matching COAST modules to mainboards. I am quite worried about creating smoke...
That's true, some manufacturers were so evil to use a custom pinout, but so far I only had issues with COAST modules on 486 boards. The module, which I showed in this video has the right ground and vcc pinout, so there was no chance to see the white smoke, but it still seems not to work properly though.
Excellent work! Videos like this make me wish I had held onto my old Megatrends HX83 socket 7 board and AMD K6 200Mhz CPU. I don't remember what happened to it when I replaced it with an Abit BH6 mainboard and a Celeron 300A that spent years overclocked to 464 Mhz.
Thank you! The only missing CPU in this test IMHO tillamook - as I remember its overclockable till 400MHz, so its interesting can it compete k6-iii+ on this clock.
Hey, you did some great tests. That brought me back so much memories.
But I've got one question: If I remember correctly from back in the days, the RAM is cached top to down, so upper memory first. If you use more memory than cachable by the chipset, the lower memory will be uncached. The i430FX should have a cachable area of 64 MByte. Thus it should run faster with only 64 MByte RAM for anything not demanding more than 64 MByte. And that is certainly true for most DOS apps and games. Greetings Peer
I think the i430 caches the lowest 64 MB, but it was Win95 that liked to allocate top-down. (And NT or Linux would allocate all over.). Regardless, your point is correct: as long as you don't _need_ the extra RAM, sticking with
The sockets on the onboard asynchro-sram seem to be longer than the chips installed.
Is there a possibility to take this up to 512KB and use it for level-3 cache with the AMD K3-500 ?
Perhaps with yet another mod on the bios ?
And what about the Pipeline Burst ? Any 512KB availablity there ?
From zero to hero. I wish I had back in the days this knowledge, as it was only about processor power and nothing else.
Quake piercing the 60FPS threshold was pretty jaw-dropping.
Great video! Thanks!
The SETMUL utility can change multiplier for AMD K6-III Plus "on the fly". On fsb = 66MHz the CPU may operate from 133MHz to 400MHz with 33MHz steps! Very awesome!
SETMUL can also be used to turn the L1 and L2 caches on and off on the fly if you need to go slower.
@@HojoNorem Exactly! The second great feature of utility.
Cache makes a difference in many cases. I remember that my K6-2 overclocked @ 600 Mhz was still slower in many things (not all) than a K6-III @ 400Mhz.
COAST modules were not usually very interchangeable with each other despite using a standardized slot, it was as if the manufacturers did not want you to be able to reuse or sell them...
Quake on K6 is not slower because of FPU alone. (Yes , it adds).
But also, because it was optimalized for Pentium instructions.
From what I understand it was extremely optimized on Assembler to the Pentium (non mmx) features.
Who knows if optimizing for K6 and even 3dnow would improve performance much.
Best wishes.
@@ajax700 It would. But 3dnow came with K6-2, not K6. And it was 1998. So ID software focused only on Q2 and later Q3, to support 3dnow!.
K6 came also after Quake was released. K5 wasn't so great, in comparation with Pentium 1, so maybe they decided they have time to optimalize engine only for one of them. Maybe they underestimated AMD, and didn't expect K6 will be so good. So choosen only Pentium. And in 1997, probably decided, they will not focus on rebuild engine to K6 too, maybe time was precious, and they were already focusing on GLQuake and Quake2.
Best wishes too.
@@warrax111 Amd almost never supplied more than around 30% of the cpu market, so it makes A LOT of sense to optimize for the most used CPUs.
They also provided compilers, tools, detailed documentation, etc.
Best wishes.
Really cool project, very interesting how far you can push this old hardware and the vrm is genius. I wonder if those coast cache modules come in different speeds and how much that would affect the system as a whole. Also would be interesting if you could find a 512k version too see if there are any benefits or downsides to that.
Keep up the great work and awesome videos!
during whole video my inner voice was screaming: K6-3! put K6-3 in there!
and than ... BOOM :D
and second thing. I remember when I was first time inspecting my socket 7 board many years ago. I was wondering what are those various slots for. and now I know :)
Just found your channel :) Very nice. I'm just getting back into old computer hardware after a break. Have to replace some plastic 30pin SIMM slots on my 386 motherboard :( lol. But found some spare slots on a 386 backplane I have that will probably never be used again due to corrosion form the battery etc.
I liked this series of videos so much! 😊
The idea of getting a piece of hardware, a motherboard, to work in a way which it only ever had the potential to do is really appealing to me. I really love the idea of extracting the most from the options of a motherboard.
I commented in the past about SMPS running at an higher frequency than the current design. Now it's probably not that interesting since the VRM seems to do its job quite well but I've done some research and there are a couple of ICs which could be used.
One is the OnSemi NCP3020B (version -B for 600MHz switching frequency) in SOIC-8 package for which there is an evaluation board, NCP3020BGEVB. Schematics, bill of material, gerber files are available. Unfortunately the board is not in stock at any of the usual distributors so I'm not sure whether it has been discontinued.
An alternative controller is TI LM2745/8 (adjustable from 50 kHz to 1 MHz) in TSSOP-14 package. It's slightly trickier to hand-solder and more expensive. It features a Power-Good open-drain output. The evaluation module, LM2745-19AEVAL, is in stock at Digi-Key in the UK for £75.372 including VAT.
Thank you very much! I'm trying to make my projects as inexpensive as possible. The parts for my VRM cost around 8-10€, of course developing the prototype was more expensive in average, since I had to try different parts and designs, but it is still by far not £75. Still thanks for suggestions.
@@necro_ware
You're welcome. The evaluation board would be a Non-Recurring Engineering (NRE) cost for the new VRM development 😉 Unfortunately they are generally expensive and they cost quite a lot more than the cost of the parts. In the industry they help speeding up development and time is money!
Even if you don't buy these evaluation boards it's good that the manufacturer of the IC makes them because they provide example of (hopefully) working schematic and layout which you can copy or use for inspiration.
@@lorenzo.c That's true! For my day time work I wouldn't go the same way, as what I do here, but here there is one absolutely free resource, which I can rely on - my time ;) Furthermore, most projects which I do are to learn something, that's why also in this case I didn't want to just copy something, but calculated everything from scratch. I really wanted to understand each click which happens in this circuit. One additional challenge which I always like to set is to be cost effective in terms of production. Of course, my engineering time is not included and that is the most expensive resource in the industry, but in the hobby the journey is the reward.
I had a hunch there should be something missing from you today,thanks for the great content !
Wow , didn't expect so big differance between K6-3+ Asynch cache, and PB cache.
K6-3+ has internal L2 cache, which is very fast and ticks at processors speed, so in this case 400mhz. L2 cache from motherboard become L3 cache, and it's used only sporadicaly, when CPU runs out of L2 cache. Seems Quake can take all CPU cache and still need more. Incredible.
yeees k6-3+ I've been waiting to see it! So nostalgic, personally. First pc I ever tore down was an IBM with a k6-3!
Wow that Cos module made a day and night difference, never thought i would see sutch a big jump, i wonder kf we can recreate them nowwdays as i have a asus board that could benifit from one :)
PS great job on the power module , I wouldnt mind picking one up at a later date :)
Must be over 25 years ago but I remember those older Asus boards with the media bus slot at the end of one of the pci slots, the p55t2p4 with the right revision I think rev 3 and above would overclock and run a bus speed of 83mhz just by setting jumpers on the board- and was the fastest board out there at the time - it was posted on Toms Hardware about 25-27 years ago but your running with a fixed clock divider! I remember it well as I bought the board for that reason, andran mine at 83Mhz up until the p2 came out without issues after I got it settled down, from memory I only run intel CPU's in the board, I had a sb16, a GUS Max and a diamond stealth 64 (S3 chipset) with an NE2000 compatible nic, all seemed to behave ok in it at that speed, I wrote a simple multi-config DOS menu that i used to assign irq and dma to the sound and scsi,I used a multi config to change settings by menu selection at boot time as needed , there were cards that didnt like the high bus speed like the AWE32 would sound wierd and I did have to swap out the 1542 adaptec scsi card for a PCI NCR one as the adaptec didnt like the overclock neither did the advansys or buslogic cards behave at overclock - great stuff it brings back great memories and always look forward to viewing your channel what you post thankyou
wonderful work.
so happy you are exploring these old parts.
Great series! Amazing to see the effect of the COAST module
Amazing! Waiting for the next step and the next improvement! You are great man!
Interesting to compare KIII+ with PII and can it run Mpeg4 video, very popular in those time)
Awesome video, as always! I'm still sad that I could not repair my Asus Socket 7 motherboard with a broken CPU socket. The main reason being that I could not find a broken MB with a good socket and I would not want to destroy a working MB.
YUS! Been waiting for this one!
That box of boards looks like a mother lode. I look forward to more very interesting videos from you in the near future.
Great job!
I asked a computer expert about that slot back in day and they said it works great but its almost impossible to find module.
Thanks for this video and the always interesting content!
I have to add something about the MMX Pentiums... they were a little more than just more cache and MMX. There was a better branch predictor, borrowed from the Pentium Pro. But that's not all! The Pentium has two pipelines, called U and V. The U pipeline could run all instructions in parallel, the V, however, could not. You carefully had to order your instructions in order to be paired on the original Pentium. This especially applied for prefixed instructions - those not available on the original 8086 - so basically several 32 bits instructions.The MMX Pentium, however, could handle more operations in the V pipeline, and hence, this, with the larger cache, and better branch predictor, have a notable impact on performance as well, as it could pair more instructions per cycle.
I wasn't aware by HOW MUCH pipelined burst cache could speed up those old CPUs... have some stick lying around and surely one or two boards - need to try it ;)
I wonder how this compares to a basic setup with celeron 266
75MHz bus was some stable overclocking everytime I tried it :)
I have the UMC burst pipe module in my m520 and it works without any issues. So I think that might be a bad module. I could test it in my PCchips m520 or if you have one you might want to try it there. Oh the m520 also has a VRM socket.
This Pentium 166 from Russia. There is an Label "ТК НИКС" - that is an big computer shop in Moscow. Exists must be for 25 years or more.
Thanks! Very interesting and... nostalgic )
Very nice benchmarking and comparison! I ought to do similar with my Socket 7/Voodoo 1 test box. It's currently running a K6-266 because that's what was in it.
A pipeline burst cache... I totally forgot about that. It had a HUGE impact. I went for a Pentium 90 instead of the P100 an invested the money into PB cache instead. This was a great decision as my home-built P90 wiped the floor with the Compaq P120 that we used in the office.
Too sad you didn't try the FSB upgrade. I totally understand the reason of course, but it would have been really interesting. Sure, not 100Mhz, that will most likely not work, but maybe 75Mhz. At least on later motherboards a moderate overclocking of the FSB to 75MHz was always stable. I also never had a PCI card that didn't work with 75Mhz. It would be interesting, if this works for the FX chipset too. Then you would get the AMD chip to 450MHz and see the impact of FSB increase.
Was kann ich noch hinzufügen, einfach perfekte Arbeit, Grüße aus Tschechien
Such beauty from a time when I was only 4 years old.
Fantastic rundown sir. I'd love to see if you could put that 64bit PCI slot to use.
The 512k COAST module gives a further performance boost. Probably a decade ago now, I tested the difference between a 256k and 512k COAST module on a P200 MMX, and there was a 5-7% performance improvement in Quake.
There's supposedly a rare 1 MB COAST module, but I never seemed to be able to find one.
a k6-III was able to work on this socket 7 motherboard??! 😱
Wow...I never knew the potential this motherboard was capable of having.
Fantastic series this was. My concern though is the build itself. Can it be put to the test with voodoo 2 (sli or not)?
What atx power supply is recommended on socket 7 motherboard btw?
Unbelievable. God damn, man, you're good :D
Amazing job!
I enjoyed the video! I would like some details though, from someone. The last CPU (printed 500MHZ) was running at 400MHZ? 66.6 MHZ FSB bus by 6 multiplier CORRECT? And it worked faster than the others because of what? The internal cashe? That external PipelineBurst cashe module really improved things.
Yes 66x6 is max possible clock in this board. K6-3+ was faster because of much faster internal L2 cache.
Great Video i love it
Thank you for the i166 and COAST once again ;)
This is defenitly the fastest S7 setup you can ever dream of.
It would be cool to see more comparsions if you’re able to test with boards that can utilize higher fsb just to see what performance difference it gives
66MHz FSB bottlenecked all CPU's starting from around 166MHz, and that bottleneck became even worse after 200MHz. MMX-Pentium's bigger L1-caches helped a little, but underlying slow 66MHz FSB and 33MHz PCI-bus strangled the whole system. I ran 350MHz K6-2 as 336/112 because it was almost as fast as 400/100. But ran cooler with stock voltages. Slightly higher PCI-bus (112/3=37.3MHz) helped too. Practically all PCI-cards I used (Matrox/Voodoo1/3Com NICs) worked with 37.3/37.5MHz PCI-speeds, sadly many began to fail with 41.5MHz PCI-bus (83MHz FSB).
I remember wondering how the FSB speed division worked on motherboards, I didn't know enough to investigate it back then though. Do you know? is it handled by a crystal or an IC somehow? I mean the division of the FSB to PCI/ISA bus speed.
From what i remember say PIIX has all the clocks supplied externally and it doesn't care where they come from. But to convert between different clock domains, it performs buffering. So there's two parts to every multi clock domain system, for one clock generation, and for other signal translation which happens on whatever logic IC that has both clock domains, and signal translation for PCI to ISA is on the PIIX. But PIIX being a PCI device, of course it's only half the picture, it never has to deal with FSB.
Let's see, i fished out a 430tx mainboard reference schematic. I can see there a dedicated clock generator IC which is part of the chipset, IMISC652 or IMISC671A. I also remember that newer boards usually use a similar multi clock domain generator IC, it's often third party, there's a bunch of suppliers, Cypress probably the main one. A tiny little thing, not noticed at a glance. So this one has a 14M318 crystal oscillator and generates PIIX clocks, 48M (bet that's the USB base clock on the PIIX), 60/66M, PCI, FSB and DRAM clock i think. I can see on this board also a 32K768 RTC crystal in either of the two places, could be on IO IC or on the PIIX. So that's it, no more crystals.
Fundamentally an AM3 system still looked the same, with that 14M318 and some cryptic proprietary clock generator spawning everything from that. AM4 looks different, there's a 25MHz crystal at the chipset; and a 48MHz one and an RTC one at the AM4 socket, and i guess all the other frequencies just come out of either of the two chips, mostly the processor just generates the clocks internally a whole bunch of them.
Fun fact: apparently fundamentally reliable, race-free clock domain transition between uncorrelated clocks isn't possible. I don't know if it's an actual fact, i haven't verified it first hand, but i've been told that it is so. But it's possible to have probability of failure astronomically low, and that's probably good enough for systems of the eras where the domains are independent. Today all interfaces are packet oriented which presents an extra opportunity, that packets are buffered input side, can be checksummed and re-sent, so you can cover up the potential unreliability of clock domain transition this way.
@@SianaGearz Now that's information! thank you very much
I wonder if a cache size increase on the onboard cache since it's detected as l3 would do anything ?
Fun fact: you can actually mod any K6-II, II+, III, or III+ so you can switch them from a K6-II up to a K6-III+
Awesome!
Have you tried to turn on write vombining on the K6-2/3 CPUs? That can help A LOT with video access and subsequently with games.
Funny misspelling :) Yeah I tried, but there are only tools which give a boost for higher resolutions, which I didn't present in this video. May be next time.
@@necro_ware Indeed :D Well, if you can turn it on for the A000-AFFF area you can see a healthy speed boost in Quake 320x200 or even in Doom (not that sure as it has a strange video memory access pattern). K6WCX surely did it for me. In a real SS7 system, I must admit, yet it should work just the same here.
Can you make similar detailed video about upgrading Slot1 440BX 98y m/b with only 2.0v & 100MHz up to Tualatin 10x100Mhz via simple PPGA Slocket?
Seems like Quake 1 is playable on all configurations. I've always wondered if K6-2+/3+ could reach a level of performance on Socket 7 to run games like Quake 3 and Unreal Tournament at at least 25 fps, with 3D acceleration of course. Seems like that would be quite a challenge, even with completely overpowered AGP cards.
Played a lot Q3 on k6-2 450 with Matrox G200 but on more modern AGP system with mvp3 chipset (chaintech 5AGM2 if I remember correctly)
Yes, Q3 is a big challenge. Q2 is fine, you can play it in 800x600 smoothly.
The issue you'll run into trying to reach those performance levels, is that these chips really need the FSB speeds of an SS7 board to get there easily. Having a faster FSB allows for much higher bandwidth between the CPU and GPU. I have yet to try it, but my FIX VA-503+ board has I think a 124Mhz option for the Front Side Bus and a whopping 1MB of onboard cache. The K6-2+ I'm using and the ram seem pretty rock solid in this configuration. It's most just a matter of finding the time to try some games. I feel like what might actually bottleneck it is the Voodoo Banshee I have in there. I wonder if anyone would know if the CPU would bottleneck the card, or if the card would bottleneck the cpu?
@@railsrust I have an MVP3 board but I only overclocked via the multiplier (5 to 5.5). Perhaps FSB is the way to go?
@@Edvinas_channel Perhaps. It depends on the silicon lottery. The ram needs to be high quality and built for 133Mhz (Pentium III era ram), and the cpu needs to be able to handle the FSB. I would make sure to pair it with a card that is actually a little too fast for the cpu like a Voodoo 3 so as to remove bottlenecks. Even this, these games really aren't what the K6-2+ and better cpus were really great at. They let you run a ton of Windows and DOS games out of one machines while making a few compromises; performance on late 90s 3D games being one. That said, it should probably be possible to run UET and Q3 at 30 or better *most* of the time, but I wouldn't expect it above 640x480. I haven't tried it yet myself, so I don't actually know. Just telling you to mute your expectations.
Did you ever actually test the media bus extension using that SCSI/SoundBlaster card?
Yes. Works fine. Initially I wanted to build a retro PC with this hardware, but as you see it turned out to be a bigger project :)