Q&A highlight
Posted8 days ago
Key Takeaways
And 2 celerons were cheaper than a CPU with double the performance?
Celerons were consumer-grade budget kit.
These were competing with PII processors in 1998, and for folks who wanted to go dual CPU it was the way to go.
There was a whole cottage industry of folks modding these CPUs as a small side hustle for people who were not comfortable with soldering onto CPU pins if you wanted to put these into a SMP system.
Performance really did mostly scale linearly with clock speed back then - but for a single CPU. The dual CPU setups were not nearly as efficient due to software not being as multi-threaded as it is today. The big win were folks with two monitors (rare!) who could run apps on their second monitor while playing games on the first. Typically you would only see frame-rate increases with CPU clock - and of course the very start of the serious 3D accelerator (3dfx, nvidia, ATI) scene back then.
When Intel switched from Slot 1 to Socket 370, there was a market for "slocket" adapters that allowed Slot 1 motherboards to take Socket 370 CPUs. The best of these adapters worked out a way to re-enable SMP on Celerons by tweaking the pin layout to disable the lock Intel had added. What made the BP6 so popular is that it was a native dual-slot Socket 370 motherboard that had this modification built in so it could use unmodified dual Celerons out of the box.
> Performance really did mostly scale linearly with clock speed back then - but for a single CPU. The dual CPU setups were not nearly as efficient due to software not being as multi-threaded as it is today. The big win were folks with two monitors (rare!) who could run apps on their second monitor while playing games on the first. Typically you would only see frame-rate increases with CPU clock - and of course the very start of the serious 3D accelerator (3dfx, nvidia, ATI) scene back then.
Even if you only had one monitor, multitasking was FAR better on a dual-CPU machine than on a single CPU system. For example, if you were extracting a ZIP file, one CPU would get pegged at 100% but the system was still responsive due to the second CPU not having any utilization. If you use a dual-Celeron BP6 system, it's a much nicer and more modern feeling experience than using a single-PII system even with the faster CPU with more cache.
I ran Slackware on my BP6 while I was in college. Of course CONFIG_SMP wasn't set in the default kernel config at the time so you had to build your own. Great for running bind, apache, sendmail, etc., and of course NetQuake servers. :)
Yes but they got hard to find in a hurry once word spread. I had two 366s in my BP6 overclocked to 550 but IIRC I had to buy a few to find two that were stable at this frequency.
> There was a whole cottage industry of folks modding these CPUs as a small side hustle for people who were not comfortable with soldering onto CPU pins if you wanted to put these into a SMP system.
No modifications necessary with the BP6.
I bought two to have one gaming machine and one coding/hacking machine (including learning about networking now that I had two computers). Geek heaven.
That would be quite the "budget" SMP build. The 366MHz "Mendocino" was based on the prior Pentium II core I believe. So quite the disparity in single-threaded workloads.
Celeron CPUs were usually CPUs that shared the same core architecture as the current Pentium standard, but often had a lower core clock speed, lower core memory speed, and/or had smaller L2 caches.
Workloads have different constraints however, and simply doubling cache, clock speed, or memory bandwidth doesn't necessarily double performance, especially when running more than one application at once. Keep in mind, this is Windows 98 /NT/2000 era here.
Symmetric multi-processing (SMP) could be of huge benefit however, far more than simple doubling any of the above factors. Running two threads at once was unheard of on the desktop. These were usually reserved for higher-binned parts, like full-fledged Pentium workstations and Xeons (usually the latter.) But Abit's board gave users a taste of that capability on a comparative budget. Were two cheaper than a single fast CPU? Probably not in all cases (depends on speeds). But Abit's board gave users an option in between a single fast Pentium and a orders of magnitude more professional workstation: A pair of cheaper CPUs for desktop SMP. And that was in reach of more people.
In short, two Celerons were probably more expensive than a single fast Pentium, but having SMP meant being able to run certain workloads faster or more workloads at once at a time when any other SMP system would have cost tons.
But in the real world, the perceived performance improvement was more than doubling. The responsiveness of your machine might seem 10 or 100x improved, because suddenly that blocking process is no longer blocking the new process you're trying to launch, or your user interface, or whatever.
It's much rarer to see misbehaving multithreaded processes hog all of the cores. Perhaps most processes are not robustly multithreaded, even in 2025. Or perhaps multithreading is a sufficiently complex engineering barrier that highly parallelized processes rarely misbehave, since they are developed to a higher standard.
Except on Windows laptops. Where, although the computer is idle, your favourite svchost.exe will heat your system and trigger thermal trottling.
This had an interesting side effect: Celerons of that era overclocked extremely well (stable 300 -> 500MHz+), due to the smaller and simpler on-die L2 cache relative to the Pentiums of the era, whose L2 cache was much larger but had to be off-die (and less amenable to overclocking) as a result.
For clock-sensitive, cache-insensitive applications, an overclocked dual Celeron could easily outperform the highest-end Pentiums of the era, especially if designed to take advantage of parallelism.
Another thing that helped the Celeron overclocking craze is Intel seemed to damage the brand badly out of the gate. The original Celerons had no cache at all, performed terribly and took a beating in PC reviews. So even though the A variants were much better this still had a stink on them.
The thing that probably helped the Celeron the most with overclocking though was they gimped them by only giving them a 66mhz front side bus speed. Since you had to increase this number to push the locked multiplier CPU speed up this was an advantage if you were going to overclock as you could buy a capable motherboard and run it at stable 100mhz. Whereas you'd have a lot more system wide problems trying to push a Pentium's 100mhz bus higher.
You could attempt to head toward ~700 but I never could keep it stable there.
> Socket 370 era Celeron processors had a Pentium II core, but Intel disabled the ability to change the multiplier to discourage overclocking
The whole point, really. Sorry, but you are.
In the Pentium / PII era, CPU speed was the product of 2 factors: front side bus speed and clock multiplier.
The original P5 and the 2nd gen P54 ran at FSB speed: 60 MHz or 66 MHz.
The later Pentiums ran at an integer multiple:
Pentium 120 = 60 MHz * 2 P 133 = 66 * 2
The PII ran at larger multipliers. I bought a PII 450 MHz on a Gigabyte motherboard that fit in my Baby-AT case, rather than the then-common ATX case.
450 = 100 MHz FSB * 4.5
That cost.
Slower PIIs had a much slower 66 MHz FSB.
PII 300 = 66 * 4.5
The Celeron had a locked multiplier but you could change the FSB.
So, take a 300 MHz chip (running at a locked 4.5x the FSB) but put it on a much faster 100 MHz FSB at the same multiplier and you got a 450 MHz chip, and because it had a much smaller but on-die L2 cache, it was more likely to be stable.
These Celerons came on a Slot 1 convertor and that needed to be modified to enable SMP operation.
Pics:
https://www.vogonswiki.com/index.php/Intel_Celeron_300A
The famous BP6 motherboard ignored that setting and forced the uniprocessor-only Celeron to run in dual processor and overclocked it as well.
So for 2 budget sub-£150 CPUs, rated for 300 MHz on a sluggish 66MHz FSB and one CPU only, you got a dual-processor 100MHz FSB machine with the raw single-core performance of something like a £500 processor.
I never had one but you bet I heard about them and strongly considered it.
People pretty routinely nearly doubled the clocks on Celeron 300As, anyway. :)
How is a small motherboard manufacturer 25 years after minor popularity in a sub-community not niche of niche?
Hacker news is IMO a niche community anyway, and I'd say the crossover between people who built their own PCs 25 years ago, during the golden years of overclocking and hacker news readers is pretty huge actually.
If you don't think this sort of article is a good fit here I don't think you are really in the target demographic anyway.
Another cool thing was that the BP6 supported Ultra DMA/66 (aka ATA/66) and it did so by adding a second controller so you had twice as many buses.
Still, I made good on my promise to never return to single core machines.
They were backwards compatible with socket 5 (you had to set the motherboard jumpers voltages though).
Some of these boards had both sdram and edo ram slots along with an agp slot, pci slots and an isa slot.
So you had an era where motherboads could take a P-75 or an amd k6 550 cpu. They could take ram scavanged from an old 486 (edo ram) or you could put in faster ram. You could run a pci grapchics card if it’s all you had (eg. A voodoo 2) or you could run an agp card. I used my old 486s isa soundblaster awe in that board for a long long time since pci was of no benefit for a soundcard.
The only set of cpus not compatible were the slot and socket 370 cpus. But they were pretty expensive anyway and it was fun to be able to frankenstein computers so much back in the day.
What I am hoping for is that this leads to a resurgence for all those used computers out there... plenty of great machines from the last decade that should have no problem being competent workstations for 90% of people's needs for the next decade onward if treated well. This is where open standards and open source truly shine.
It wasn't much but I could run Alice, Max Payne, GTA 3, Dungeon Siege on there, all at like mid settings, so I was a pretty happy camper for a high school kid putting paper route money into my own PC.
I think I would have done the same with my AWE64 Gold if that was still an option for me in the early 2000s.
I definitely used the Sound Blaster with my 486DX100, and I recalled migrating it to at least one other machine after that; it was nice for the joystick port and also the better wavetable synth on classic games.
>along with an agp slot
Non working AGP slot, or rather working until you tried to play 3d games with 3D accelerator actually using AGP features, then you got crashes no matter the chipset (VIA, ALI). Solution was switching to x1 mode, disabling sideband signaling or just swapping to a 3dfx card.
>amd k6 550 cpu
thats year 2000
>The only set of cpus not compatible were the slot and socket 370 cpus. But they were pretty expensive anyway
You are comparing bottom of the barrel AMD CPUs with top spec Pentium 3s. Correct comparison should be against Celerons. January 2000 prices https://akiba-pc.watch.impress.co.jp/hotline/20000617/p_cpu....
K6-III/450 14,550 $140
K6-III/400 8,980 $85
Celeron 300A $57
300A@450MHz beats K6-III/450@550MHz in every possible benchmark.
by June 17 2000 https://akiba-pc.watch.impress.co.jp/hotline/20000617/p_cpu....
Celeron 533A 10,570 $100
Celeron 366MHz 7,700 $73
Duron 600MHz 9,990 $95
K6-III/450 24,800 $236 !??!!?
K6-III/400 14,800 $140
K6-2/550 7,949 $76
K6-2/533 5,970 $57
K6-2/500 5,350 $50
$76 K6-2/550 is slower than $73 Celeron 366, not to mention pulverized in benchmarks if you happened to find Celeron capable of 100MHz fsb.
Old slow ram makes K6 setup even slower. You would think the benefit were much cheaper motherboards, but even that wasnt the case. SS7 boards started at ~$75 while Abit BE6-2 was $90 and cheapest 440BX ones (P2XBL) $65. K6-2/550 3DNow! (100MHz Bus) $90 vs Celeron 500 $93 https://archive.org/details/computer-shopper-2000-07/page/n3...
Slot1 made much more sense, only release of K7 made AMD competitive again with Duron on the low end and Athlon way ahead of P3.
EDO was from the 2nd generation of Pentium. It was not a 486 thing and I never even heard of any 486 that could use EDO.
The 1st gen Pentium chips were 5V parts running at 60 MHz or 66 MHz.
The 2nd gen were 3.3V and ran at 75, 90 or 100 MHz.
The critical development for EDO was Intel's Triton chipset (the 82430FX) which added EDO support for about 15% more memory bandwidth.
I ran the testing labs for PC Pro magazine at that time and Tulip Computers of the Netherlands sent in a Pentium with an SIS chipset that could detect EDO but not use EDO's timings, so it was no quicker. I wrote about it.
(This was circa 1995. 486 introduced 1989. P5 (5V) 1993. P54C (3.3V) 1994.)
Tulip threatened to sue. The board and some lawyers flew to London. I demonstrated that their PC could detect EDO and showed a message that EDO was fitted which did not appear if FP-mode DRAM was.
I then showed that a machine with a 430FX chipset was circa 15% quicker with EDO than FP RAM, and the Tulip was the same speed.
The threat of litigation was withdrawn and Tulip used our phone to make an international call to Taiwan there and then to shout at SIS.
I worked extensively with this stuff.
No 486s with EDO to the best of my expert knowledge.
Legit 486 EDO chipset is UMC 8881E/F https://www.vogons.org/viewtopic.php?t=93147 "486 board with UMC 8881E/8886B: The winner is: EDO without L2 (if your only other option is L2 at 3-2-2-2)"
TLDR: standard on any 486 motherboard L2 cache masks any EDO mode gains, even when EDO runs with fastest supported 3-1-1-1 timings.
Yes, that is what I discovered, experimentally verified, and published -- in 1995 or so, some 24 years before that post.
> Legit 486 EDO chipset is UMC 8881E/F
I am not sure what you're saying here. This seems like a very messy mix of conclusions...
* At least one 486 chipset can use EDO -- sure, no problem, I can believe that. But was it designed to use EDO timings? I doubt that: 486 family chips were fading legacy tech when EDO was first invented.
* Quake benchmarks: Quake was hand-coded to interleave FPU and integer ops in a way that used an inherent property of the Pentium hardware design. I went into this in some depth in 2016 here:
https://liam-on-linux.livejournal.com/49259.html
The int/FP interleaving didn't work on any other CPU. Not on 486, 5x86, 6x86, anything. So small performance differences on a 486-class chip do not really tell us anything meaningful here, IMHO.
Yes, this chipset can do EDO 3-1-1-1. Of course problem is Cache can also run at 3-2-2-2 or 3-1-1-1 and board doesnt have buffers between ram and rest of the system thus in EDO mode you need one more cycle to close the page after every burst to unlock data bus. EDO = extended Output = ram keeps outputting last accessed data.
Early Intel Pentium chipsets included additional chips performing that buffering between system and ram (for example 82438VX). Still even 430VX slowed down to standard FPM 7-3-3-3 timings with EDO and Async cache, only burst pipelined cache allowed 6-2-2-2.
The big Pentium era performance jump seemed to come from pipelined cache significantly unlocking ram subsystem potential https://dependency-injection.com/early-pentium-chipsets/ Just the EDO vs FPM difference was small around 1-3% depending on the cache situation.
All right, I will take your word for it. When did it come out, though?
I think EDO RAM didn't start to appear until the early/mid 1990s -- 1993/1994 or so at the earliest. This was very late in the lifespan of 486 chips (and enhanced ones like 586s) and so this would have been a small performance tweak for very low-end budget hardware, surely?
> Early Intel Pentium chipsets included additional chips performing that buffering between system and ram (for example 82438VX).
Hang on. The 430VX was not an early Pentium chipset. It was a late one.
It launched in 1996: https://theretroweb.com/chipsets/277
After the "Triton", the 430FX:
https://dosdays.co.uk/topics/intel_chipsets.php
The dominant early Intel chipset for Pentium hardware was the 430NX "Neptune". I had a Neptune-based PC at work, originally with a Pentium 66 in it, later replaced with a PODP, the Pentium Overdrive, with a clock-doubled 3.3V P54 chip in a socket adaptor.
Neptune was nothing special and had no performance boosts to speak of. The only interesting thing is that as it didn't have built-in EIDE, it was often on a PCI card. I removed it from my all-SCSI machine for a "purer" setup with nothing using the EIDE I/O ports and DMA channels.
Triton (430FX) brought in EDO support, and was as I said about 15% faster with all other factors being equal: same CPU, same cache, same drives, same graphics, etc.
This is the time period when I developed the 32-bit version of PC Pro magazine's 16-bit Windows benchmark. I was very familiar with PC performance and components back then.
The DosDays website is confusing because it lists the chipsets in this order:
NX "Neptune"
FX "Triton"
HX "Triton II"
TX "Triton IV"
VX "Triton III"
... when its own dates show that the TX came later, and it really went:
NX
FX
HX/VX more or less simultaneously and both termed Triton II
TX <- I don't think I ever saw this
I strongly disagree with your comment about cache.
Cache/no cache was huge. Write-back vs write-through was huge. Pipelined burst cache helped a lot but any L2 cache was good.
No, the type of cache wasn't a big difference: having it at all was what mattered. Cheap cloners had no L2 cache and modified the startup messages to say "writeback cache" meaning that only L1 cache was present.
Aside from bargain-basement skipware, most Pentium boxes that were any good had Intel chipsets and L2 cache. Usually only enough for caching the first 64MB.
late, and only in 4th revision of this particular chipset :) Its the exception to the rule.
>Hang on. The 430VX was not an early Pentium chipset. It was a late one. After the "Triton", the 430FX:
FX had similar buffers - 82438FX
> early Intel chipset for Pentium hardware was the 430NX "Neptune"
2x 82433NX LBX "data path between the host CPU/Cache and main memory", no mention of FPM/EDO in datasheet and only goes down to standard FPM X-3-3-3 timings. Afaik still works with EDO just with no speed difference.
>I strongly disagree with your comment about cache. Cache/no cache was huge.
Only on chipsets not supporting faster EDO timings. No L2 is not a big deal with EDO because EDO timings are already almost as fast as Async L2 cache (7-2-2-2 vs 3-2-2-2). No L2+EDO https://dependency-injection.com/intel-430fx-triton-l2-cache... 2% slower in Doom but 8% faster in Quake. Comparison between Async L2 + FPM vs no L2 + EDO would look even better for EDO.
PB cache on the other hand was an easy 10-20% bump over Async L2.
Now take a chipset that utterly fails with EDO like sis 501 and difference is indeed dramatic https://dependency-injection.com/2mb-cache-benchmarks/
>No, the type of cache wasn't a big difference
I literally linked tests that show otherwise :) Intel datasheets explain why, PB not only allowed 3-1-1-1 cache timings but also unlocked faster ram modes.
>modified the startup messages to say "writeback cache"
Wasnt PCchips pretty much the only fraud that made boards with no L2 cache, with some other vendors (amptron, kaimei, jamicon) selling same relabeled pcchips? There was a funny case of Octek selling some models plastered with "Dynamic Cache Architecture" stickers while the cache was build into special EDRAM ram, big problem being most of those models shipped with chipset unable to support said EDRAM :) example https://theretroweb.com/motherboards/s/octek-hippo-vl-2
I recall that there was a while during the Athlon 64 era that DFI was the gaming board to get. But I feel like I hear references to Abit more often than DFI.
I think my old Opteron machine with a DFI board is kicking around somewhere still.
https://www.neilvandyke.org/cheap-pc-2000/
That page includes pricing info for each component, and how I bought it. For example:
> Abit BP6 Dual PPGA Socket-370 motherboard, UDMA-66, 2 ISA, 5 PCI, AGP 2X, 3 168-pin PC100 ECC, max. 1GB RAM. Retail version. (Essential Computing $120 + $14.25 UPS Ground + $3.60 insurance = $137.95)
> Intel Celeron 500 Retail version, with warranty and CPU fan and heat sink. (Egghead $135.99 + free UPS Ground = $135.99)
The box was my workstation, and for a time also a public Web server on ADSL. I never actually added a second Celeron (cost money, and I still wasn't feeling CPU pressure) nor the UDMA-66 (reported to be less reliable).
Like this?
https://upload.wikimedia.org/wikipedia/commons/7/7b/Abit_dua...
As impressive as they look, and as effective as they were at the time, I believe even a modern stock Intel cooler would be more effective; CPUs of the time only had TDPs around 20-30W, in comparison to the 100W+ they are now at.
You are 100% correct, how could I forget? :(
I think I got it mixed up because I bought a Zalman GPU cooler which had similar design for my GPU some years after.
Thanks for the trip down memory lane.
The coolermaster cpu sink i'm using now is big, but not particularly pretty.
0: https://toffeex.com/heat-exchangers , https://toffeex.com/heat-sinks
In fall of 1999, I built my first PC with an Abit BE6 to use with an Intel Pentium 3 'Coppermine' 500 MHz. I was a fifteen year old kid working at pizzeria in the midwest making minimum wage to feed my computer hobby. At the time I was reading hardocp, and compiled a list of "good" BX motherboards to try and find at computer show that was organized on a semi-regular basis at the local fairgrounds. This event saw numerous mom&pop computer stores travel from hours away to sell custom PCs, software, and hardware. I remember having a bit of buyer's remorse because I actually wanted the BE6-II which featured the ability to change the front side bus in 1 MHz steps, while the older BE6, only had a set of fixed multiplers and PCI dividers. My 500 MHz coppemine (5 x 100 MHz) didn't post at 750 MHz (5 x 150 MHz) and was unstable above 667 MHz (5 x 133 MHz). This overclock still 'saved' me a considerable amount of money by allowing me to purchase the cheapest part and squeeze performance out of it. That computer hobby led me down a path of studying computer engineering and my eventual departure (escape) from the Midwest.
Years later, in 2015 I moved to Taipei, and remember walking around Neihu seeing all the headquarters of the computer part manufacturers I used in my childhood (Liteon, ECS, Nvidia). I didn't realize that Abit's former headquarters on 陽光街 is right next to many of the places I've been living and working next to for the past decade.
Another memory from that time was buying 128 MB of SD-RAM from Crucial (Micron). I remember being a little pissed because the price had gone up 50% due to the 921 earthquake, which killed thousands and left many homeless, and knocked the fabs offline which led to a supply shortage.
It was a good time when IRC, AIM, and physical electronics shopping was still a thing. The only big tech presence that techies hated was Microsoft. Sigh.
Mine was not very stable under even moderate overclocking though!
Good times!
Besides being able to overclock anything to its absolute limits, they have repurposed a keyboard microcontroller to monitor temperatures, fans and voltages on the board (aka uGuru), providing unparalleled flexibility and reliability when no other company was able to provide at that time, and after them.
Setting fan curves for temperature response was great and allowed me to run AthlonXP at 2200Mhz (200x11@1T) without excessive noise for normal tasks.
Their AN8 Ultra provided a rock solid foundation for my AMD64 system, too.
None of my capacitors have leaked/bulged despite using systems under relatively heavy load.
I like they show schematics in their materials and still have a sticker from an old celeron build. I booted it up recently and it still works.
Good memories.
I remember being surprised that HP did not make the boards themselves.
I recall they had one BIOS release that bricked boards, which I missed by not downloading it the first day of availability. The board was eventually destroyed by a sketchy PSU which killed the AGP slot and the Geforce 6800LE that was in it.
Later I tried an IP35-E, but had problems (two different memory sets failed) so I returned it for the Gigabyte P35-DS3L I actually wanted but was out of stock at the time.
Not affiliated with Hacker News or Y Combinator. We simply enrich the public API with analytics.