mkarcher wrote on 2025-04-23, 15:59:
shamino wrote on 2025-04-03, 10:22:
DDR has greater setup/latency penalties than regular SDRAM. DDR doesn't fully double the performance but doubling the bus width does.
"citation required". Comparing DDR and SDR at the same clock frequency, the setup/latency penalties are expected to be identical. In the case of the GeForce 2MX, we are looking at a memory clock of 166MHz, i.e. PC166 / DDR333. Looking at a random 166MHz SDR 8Mx32 chip, in this case the IS42S32800J-6BL , I find them to be specified for CL3 at 166MHz. Looking at a similar DDR chip, I found the IS43R32800D-6BL, which is specified for CL2.5 at 166MHz. This is actually the same latency. While the DDR RAM is half a clock cycle faster to output the first 32 bits, the second 32 bits appear half a clock later, which is exactly the same point in time the whole 64 bits by two SDR chips appear. If you need the whole 64 bits at once, DDR and SDR in this example (I just picked 8Mx32 chips I can buy right now at mouser, without any cherry picking for specs) are equally fast. If you can profit from the first 32 bits being early, 32-bit DDR is actually faster than 64-bit SDR.
You might be right about this. I went over to the "VGA Legacy" site and looked up some RAM chips that were actually used on Geforce2 MX cards.
An Asus V7100 "Deluxe Combo" which I think has very ordinary SDRAM uses 128-bit worth of
Samsung K4S643232E-TC60
and an MSI MS-8817 is photographed with
Samsung K4S643232C-TC60 which I'm guessing is just an older version of the same RAM
These are presumably clocked at 166MHz CL3 on these cards
The oddball Creative 64-bit DDR card uses
Hyundai HY5DV651622 TC-G7
which is only rated for 143MHz, so there's the obvious explanation right there.
But let's suppose Creative had actually used the 6ns rated version of those DDR chips, matching what 128-bit SDR cards were using.
I looked up datasheets for the above RAM chips and when compared at 6ns, I don't see any glaring difference in their latencies. Whatever small differences there are, I don't know if they have any significance.
The figures for tRRD, tRCD, tRP all match.
The HY5DV651622 at 6ns (if Creative had used that grade, as they should have) can be used at 166MHz at either CL2 or CL3. I'm not clear on the consequences of this choice, but the text of the datasheet implies that using CL3 gives it a longer "pipeline", so I guess it can do more consecutive reads at that latency. It seems like video cards always prefer higher latencies to get this advantage, or whatever the advantage is, so I'm going to guess that if using this RAM, the card would probably run the RAM at 166MHz (DDR333) CL3.
Comparing the datasheets for the above memory ICs (which appeared on real cards at the time), but upgrading Creative's 7ns DDR to the 6ns rated chips, it looks like the latencies are very equal. The only disadvantage that the DDR might have is that because of the narrower 64-bit bus width, maybe the DDR setup would spend a greater percentage of it's time dealing with latencies in between those 64-bit reads, including maybe needing to switch Rows more often. But I don't know if that's actually the case (how big is a row? This stuff is too esoteric for me). The duration of those latencies when they do occur appears to be the same as the SDR chips, it's just a question of whether they'd be invoked equally often.
It would be interesting to overclock a Creative 64-bit DDR Geforce2 MX to run the RAM at 166MHz and see how closely it matches the performance of a conventional 128-bit SDR card at the same clock. The downgraded clock speed is probably most, if not all of the difference that makes that card slower.
