It sounds like a problem related to memory interleaving. He doesn't say whether the memory modules are identical. Could also be a poor performing motherboard.
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.
> 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.
This reminds me that on an Amiga 600 or 1200, if you add more than 4 (IIRC?) megabytes of RAM through usual means, the PCMCIA slot becomes unusable due to addressing conflicts.
There are workarounds, of course. For instance, the A1208 expansion has a jumper that limits added memory from 8MB to 4MB explicitly so that PCMCIA can be used.
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.
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.
Also, internal CPU caches grow over time - in 286 and before, just was not any cache; in 386 first included page cache for mmu - stores tables with mostly used pages; in next generations sometimes advertised grow of page cache.
So yes, even when your cpu could address similar size of ram, possible it don't have enough page cache for your application.
how funny is this, we used to spend weeks fitting assets into 4MB, and now we spend weeks trying to figure out why a 'Hello World' microservice is OOM-ing in a container with 2GB.
We traded the 'Mo RAM' for 'Mo Layers,' and in the process, we lost the ability to reason about what the hardware is actually doing. Sanglard’s breakdowns are always a sobering cold shower for those of us pampered by modern GC and JITs
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!
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.
Reminds me a bit about installing one of my 128MB 72 pin SIMM modules in a 486, it has the same caching issues. Most board will not accept them anyway (I have both FP and Edo ones) but if you put a lower capacity one in the first slot they will happily boot and accept the full ram amount if all lanes are occupied (which is not a given on all 486 motherboards). Also remember to enable quick ram check or you will be getting more coffee.
Does the language in this not make sense to anyone? Is it trying to say the the L2 cache provided by the chipset is not able to access memory past a specific address?
On the 486 the L2 tag memory is a separate external chip, as are the L2 chips. Why waste space on physical address bits that will never be used? DOS is never using that much memory. So only low addresses are cached.
It sounds like a problem related to memory interleaving. He doesn't say whether the memory modules are identical. Could also be a poor performing motherboard.
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.
> 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.
This reminds me that on an Amiga 600 or 1200, if you add more than 4 (IIRC?) megabytes of RAM through usual means, the PCMCIA slot becomes unusable due to addressing conflicts.
There are workarounds, of course. For instance, the A1208 expansion has a jumper that limits added memory from 8MB to 4MB explicitly so that PCMCIA can be used.
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.
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.
That was a PPC603/604 limitation if you wanted multi CPU’s.
Also, internal CPU caches grow over time - in 286 and before, just was not any cache; in 386 first included page cache for mmu - stores tables with mostly used pages; in next generations sometimes advertised grow of page cache.
So yes, even when your cpu could address similar size of ram, possible it don't have enough page cache for your application.
how funny is this, we used to spend weeks fitting assets into 4MB, and now we spend weeks trying to figure out why a 'Hello World' microservice is OOM-ing in a container with 2GB.
We traded the 'Mo RAM' for 'Mo Layers,' and in the process, we lost the ability to reason about what the hardware is actually doing. Sanglard’s breakdowns are always a sobering cold shower for those of us pampered by modern GC and JITs
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?
That would be around $7,900 USD.
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!
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.
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.
Reminds me a bit about installing one of my 128MB 72 pin SIMM modules in a 486, it has the same caching issues. Most board will not accept them anyway (I have both FP and Edo ones) but if you put a lower capacity one in the first slot they will happily boot and accept the full ram amount if all lanes are occupied (which is not a given on all 486 motherboards). Also remember to enable quick ram check or you will be getting more coffee.
Does the language in this not make sense to anyone? Is it trying to say the the L2 cache provided by the chipset is not able to access memory past a specific address?
On the 486 the L2 tag memory is a separate external chip, as are the L2 chips. Why waste space on physical address bits that will never be used? DOS is never using that much memory. So only low addresses are cached.
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?
Some motherboards supported larger tag ram chips, but not all.
Some chipsets allowed that, but not all
Rip Rudy.