Mo RAM, Mo Problems (2025)

Google says sdram in 1997 was 7 to 10 dollars per megabyte. So 384 would be 3840 not 40,000 am I missing something here?

Around the turn of the millennium I had a Sony Vaio 505TX, which had the same chipset. My machine was running Linux, and I maxed it out to 128MB RAM.

There was a kernel patch for this chipset back then, which treated all memory above the lower 64MB as a RAM disk, which could then be used as swap space.

This prioritized the faster portion of RAM while still having very fast swapping.

This would have also still been true even roughly a decade later, during which time the industry was going through a transition from 32-bit computing to 64-bit, and large amounts of RAM read from BIOS in pre-UEFI systems were slower to boot the more memory you had!

Imagine young would-become engineers at the time finding that adding that second stick to their laptop did in fact, not make their systems magically faster.

It is funny to see how these older machines perform at their higher end limits. I'm guessing the idea on this was that if you needed that much RAM, the sacrifice of L2 cache was a worth while trade off.

It was only a few weeks ago that I found out the original BeBOX computers would switch off L2 cache when running in dual CPU mode. It was just a limitation of the memory controller. Again, the thinking of, if you need the extra compute over memory bus it would be a worth while trade off.

Many modern apps seem to cache based on total ram installed, and don’t seem to scale well to larger than normal systems. Chrome, I’m looking at you.

My 1997 mainboard had extensible tag-ram, if I remember correctly. Perhaps this is the issue?

Google says sdram in 1997 was 7 to 10 dollars per megabyte. So 384 would be 3840 not 40,000 am I missing something here?

Possibly inflation adjusted?

I had a desktop PC that I bought (as a pile of bits!) with 512MB of RAM in 1999 and I sure as hell didn't pay more than a couple of hundred for memory. That might have been EDO rather than SDRAM though but I can't see the price difference being that much!

Many modern apps seem to cache based on total ram installed, and don’t seem to scale well to larger than normal systems. Chrome, I’m looking at you.

Imagine young would-become engineers at the time finding that adding that second stick to their laptop did in fact, not make their systems magically faster.

Honestly asking though is it worth that trade off? I enjoy watching people benchmark older Intel x86 based chips and without cache they are frankly awful slow. I'm not sure two without cache beat one with. The BeBox did run a totally different processor though so I have zero domain knowledge for that which is why I'm genuinely curious.

Looks like the BeBox motherboard didn't have the external L2 in the first place.

Besides web sources, logic dictates this as well: Since dual-cpu was its selling point, it wouldn't make sense to ship a disabled L2 implementaton on the mobo at extra cost. There was no single-cpu model.

Around the turn of the millennium I had a Sony Vaio 505TX, which had the same chipset. My machine was running Linux, and I maxed it out to 128MB RAM.

There was a kernel patch for this chipset back then, which treated all memory above the lower 64MB as a RAM disk, which could then be used as swap space.

This prioritized the faster portion of RAM while still having very fast swapping.

In the modern era we'd probably repurpose NUMA support if this issue came up again, so that tasks would prioritize the fast portion of memory but the remainder would be fully usable as RAM (with fewer of the extra copies you'd have from "swap" use).

Too late to edit - I just saw that the Vaio in fact had the 430TX chipset, not the 430FX. Both were artificially capped at 64MB of fast RAM, as they were late Pentium chipsets, and Intel rather wanted to sell the then-new Pentium II chips and chipsets if you wanted to have more memory.

My 1997 mainboard had extensible tag-ram, if I remember correctly. Perhaps this is the issue?

Some motherboards supported larger tag ram chips, but not all.

Some chipsets allowed that, but not all

Possibly inflation adjusted?

That would be around $7,900 USD.

My computer had 16Mb in 1997, and it was lower-range but not the absolute bottom.

It looks like Anandtech listed 128Mb for $300 (not inflation adjusted) in 1997. It fell to $150 in 1998 and by 1999 you could buy it for $100.

So 512Mb RAM by the end of 1999 for ~$200 was plausible.

> Both were artificially capped at 64MB of fast RAM, as they were late Pentium chipsets, and Intel rather wanted to sell the then-new Pentium II chips and chipsets if you wanted to have more memory.

Intel being Intel, back then and now.

Looks like the BeBox motherboard didn't have the external L2 in the first place.

Some motherboards supported larger tag ram chips, but not all.

Some chipsets allowed that, but not all

That was a PPC603/604 limitation if you wanted multi CPU’s.

That would be around $7,900 USD.

My computer had 16Mb in 1997, and it was lower-range but not the absolute bottom.

It looks like Anandtech listed 128Mb for $300 (not inflation adjusted) in 1997. It fell to $150 in 1998 and by 1999 you could buy it for $100.

So 512Mb RAM by the end of 1999 for ~$200 was plausible.

That was a PPC603/604 limitation if you wanted multi CPU’s.

> Both were artificially capped at 64MB of fast RAM, as they were late Pentium chipsets, and Intel rather wanted to sell the then-new Pentium II chips and chipsets if you wanted to have more memory.

Intel being Intel, back then and now.

Feb 16, 2025

Mo RAM, mo problems

As a retro-computer enthusiast, it seems that parts are either insanely expensive or dirt cheap. If the first case has obvious problems, the second can also lead to issues.

When I built the Quake PC, the motherboard and HDD were worth their weight in gold. But the price of RAM modules was ridiculously low. So I maxed out by buying $40,000 worth of 1997 SDRAM, namely 384 MiB, for the price of $60.

From 44 fps to 33 fps

After I got the machine working, I ran benchmarks for weeks. I was constantly swapping video-cards, changing RAM types (SDRAM, EDO), adding RAM, removing RAM, and testing different CPUs. The CPU in my collection that ran Quake the best was the Pentium MM 233MHz clocking demo1 at 44.6 fps. That figure was consistent with benchmarks of the era.

I wrote an article about winquake then took a break from 1997. A month later I had the idea to measure Michael Abrash assembly optimizations. I ran the same benchmark again. But this time I measured 33 fps. That was nearly 25% slower. What happened?

Troubleshooting

I tried pretty much everything I could think of. I swapped the graphic card, removed all the 3D accelerators, updated the drivers, downgraded the drivers, wiped the whole system, re-installed everything, double checked that I was still using the MMX 233Mhz, and verified the frequency multiplier. Still 33 fps.

Did a RAM module go bad? I tried to remove one of them. 33 fps. Remove a second one (leaving only one). Now the game ran at 44 fps. Two modules going bad? Hm, that seems weird. I tried to swap modules, leaving only one in the machine. All of them ran the game a 44 fps. Only when two of more are in the machine, the framerate drops back to 33 fps.

Epiphany

I vaguely remembered something about having too much RAM from the excellent Upgrading and Repairing PCs 10th edition.

That last issue is one that many people are not aware of. The 430FX chipset can only cache up to 64M of main memory. This means that if you install more than 64M of RAM in your system, performance will suffer greatly.

Now, many think this won’t be that much of a problem—after all, they don't normally run enough software to load past the first 64M anyway. That is another misunderstanding, because Windows 95 and NT load from the top down.

This means, for example, that if you install 96M of RAM (one 64M and one 32M bank), then virtually all of your software, including the main operating system, will be loading into the non-cached region above 64M. Only if you use more than 32M would you begin to see an improvement in performance. tem.

- Upgrading and Repairing PCs 10th edition

My motherboard, an XA100, is from 1998 and advertised[1] as supporting caching of 512MiB, but there is clearly something amiss there. It seems it actually can only cache in the neighborhood of 128 MiB. Which means any amount over that, made everything run without a L2 cache!

And that is the story of how I made that PC faster, by removing RAM.

References

Hacker Times

Hacker Times

Mo RAM, Mo Problems (2025)

Discussion

Discussion