Tech Flashback: My Collection of SDRAM – 128Mb & ECC Modules

Synchronous Dynamic Random Access Memory, or SDRAM for short, was the JEDEC standardized successor to the asynchronous DRAM used in SIMMs. Demonstrated in 1993, and the de-facto type of memory by 2000, it was employed by later Pentium I systems through to the earliest Pentium 4 motherboards. Module sizes ranged from about 32Mb to 512Mb and operating clock speeds ranged from 66Mhz to 133Mhz normally, with ICs rated for 143Mhz and overclocking modules up to 150Mhz formerly available.

Without going too deep into the technicalities, SIMMs asynchronous operation can be likened to putting in an address, waiting for a certain amount of time, and reading the result. SDRAM featured more intelligence, with each RAM chip having its own state machine, synchronized by an external clock resulting in data being accepted or appearing on the output with defined latency timings for operations and the ability to perform burst operations.

While the technicalities of the RAM are rarely important for end users, SDRAM was a little different than the SIMMs it replaced, and bought a multitude of benefits to the user. SDRAM modules are larger than the modules which they replaced (168-pin), with smaller pitched contacts on both sides of the module independent (DIMM) rather than redundant. SDRAM had a wider 64-bit data bus (72-bit in case of ECC) allowing for better bandwidth and an elimination of “modules must be installed in pairs or quads” rules. SDRAM with its synchronous nature evolved to increase clock speeds which further pushed performance. Additionally, it improved upon the limited “jumper wire” style presence detect of SIMMs with a standardized serial EEPROM which contained Serial Presence Detect (SPD) data describing the module’s size, latency and manufacturing parameters, allowing motherboards to automatically configure themselves for best performance (and in some cases, also introduced incompatibility issues).

While I generally look back fondly on SDRAM as one of the important stepping stones towards DDR SDRAM and newer technologies, problems with incompatibilities as speeds increased did occur partly due to incorrect interpretation of SPD data, and due to the proliferation of generic low-cost modules with incorrectly rated ICs for the advertised SPD timings or poor PCB designs which caused corruption. As a result, Intel had to step in and standardize the PC66, PC100 and PC133 speeds at the module level and this helped to improve the compatibility and reliability of SDRAM as a whole.

There were also incompatibilities that arose due to motherboard chipset limitations as to how many banks of RAM they could drive, often intermixed with vague language such as “double sided” or “single sided” as a way of trying to convey this concept in a way that your average consumer might understand, but ignoring the technicalities behind the issue.

Later, towards the end of SDRAM’s life, compatibility between modules made with chips of high density also caused incompatibility, but by that time, SDRAM was much less important as it was superseded by DDR SDRAM.

In this post (and the next), I will catalogue my collection of remaining SDRAM. Unfortunately, due to an excess of RAM and a need to clear out some room, I regrettably disposed of many 64Mb and 128Mb PC66 and PC100 modules, owing to the larger number of 256Mb and 512Mb PC133 modules I had in my possession. As a result, my collection represents the later days of the SDRAM technology. Sadly, I didn’t even save my first set of SDRAM – a PC66 set of two 32Mb sticks made with LG Electronics memory if I recall correctly.

128 Megabyte Modules

The first set of modules from my collection are all 128Mb modules. These modules were the smallest ones I kept as I had much larger modules available. A quick hint for those with lots of memory modules of the same length lying around – SDRAM has two notches, and all of the DDR variants have one, making them easy to spot.

Century (?) DTV8A8HT7F3-JC

20160117-1108-2129 20160117-1108-2130

This module is a bit of a delight as it’s an all Japanese module. The RAM chips are Hitachi Japan 5264805FTT75 dated Week 38 of 2001. Note the part number ending in 75 indicating 7.5ns timings or 133Mhz in other words. The actual manufacturer is hard to tell, but Century is written on the PCB. It features nice resistor packs on most of the lines, and distributed ceramic caps in-between each chip and seems to be a quality make.

Generic Unbranded Module (PC-100)

20160117-1108-2127 20160117-1108-2128

This one appears to be an unknown generic module, marked WT59427 with a C-tick number of N298. The chips all come from Hyundai Electronics Korea, dated Week 6 of 2000, of type GM72V66841XT7J making this a PC100 memory of 2Mbits x 8 bits x 4 banks configuration. This one doesn’t have any resistor packs on the data lines but does have distributed capacitors.

Legend L1664PC3

20160117-1114-2156 20160117-1114-2157

This module was branded Legend, an Australian memory brand, and is made using Hyundai Electronics chips from Korea, part number HY57V28820AT-H which appears to be a 4Mbits x 8 bits x 4 banks. This is a later Hyundai product, where the last letter after the hyphen indicates speed (i.e. 6 = 166Mhz, K & H = 133Mhz, 8 = 125Mhz, P&S = 100Mhz). There are no distributed MLCC caps on this board, but the resistor packs are still there. We can see that this module claims to be Intel Certified, which again, is a reference to the Intel PC133 standard which aimed to fix SDRAM compatibility issues. This uses higher capacity chips than the one above, as it has half the number of chips and free pads on the rear, so the same PCB could be potentially used to make a 256Mb module. It seems likely that this module was supplied in March of 2001, with the chips marked Week 7 of 2001.

It definitely deserves a mention, in case readers were not aware, that Hyundai was undergoing a transition at the time into becoming Hynix Semiconductors (in 2001), which today, is known as SK Hynix.

Legend L1664P33

20160117-1117-2168 20160117-1117-2169

This was another one from legend, but without Intel certification. What’s changed? The board is a little shorter, without the resistor packs on the data lines, and the Hyundai Electronics memory chips appear to be an older, lower-density part. The part number is GM72V66841ET75, which appears to be a 2Mbits x 8 bits x 4 banks configuration, suited for 133Mhz usage. The designations are 7 for 143Mhz, 75 for 133Mhz, 8 for 125Mhz, 7K for PC100 at 2-2-2, and 7J for PC100 at 3-2-2 latencies. The chips appear to be marked Week 39 of 2000.

Nova N1664P63-081LS732

20160117-1113-2152 20160117-1113-2153

As a brand, Nova is relatively unknown. But when they’re placed side by side with the Legend stuff, it’s clear from the labels that Legend has something to do with Nova. Some forum postings seem to claim that Nova products aren’t particularly good, but it’s rather hard to tell what it’s really made of when the chips themselves have been re-marked with Nova 0140168M2 – a rather meaningless designation with no datasheet available. The batch codes for each and every chip is different in the corner, suggesting that this may be lower-grade “component recovery” style operations, but built onto their own PCB design. At least the board itself has a few scattered capacitors and resistor packs, and it seems it could accommodate more chips for a larger module as well.

However, there seems to be one dead giveaway as to who makes the chips – notice how the chips have little “bites” taken out of each side with the metal substrate frame poking out? This seems to be a characteristic of Micron Technology based products … oh … here comes one now …

Micron MT8LSDT1664AG-133B1

20160117-1115-2160 20160117-1115-2161

This is a Micron Technology OEM module, and judging from the part number of 38L3915, it seems this was pulled from an IBM machine at one point in time. Looking closely at the chips, you can see the batch code markings in the corner and the same bite marks as the Nova module. The chips on this module are marked 48LC16M8A2-75 in a 4Mbits x 8 bits x 4 banks configuration and the 75 designation indicates 7.5ns/133Mhz operation.

This module is a bit taller than the others, with more generous spacing, and it seems set up to accommodate ECC configuration and modules of twice the size. The PCB has PCSDRAM REV1.0 etched into it, which is not the only module to do so as you will see later. The PCB appears to be made in Week 4 of 2001.

Hynix HYM71V16635HCT8-H

20160117-1117-2170 20160117-1117-2171

Of course, Hynix was in on the OEM market as well, this one dated Week 36 of 2001 is on a darker PCB branded Hynix Korea, using HY57V28820HCT-H ICs very similarly to the “Intel approved” Legend memory module. From the label on the rear, this module was pulled from a Compaq machine, and the PCB appears designed for possible ECC usage but not double-sided expansion.

NCP NC2996

20160117-1115-2162 20160117-1115-2163

Another brand of RAM we would see quite often is NCP. Lots of people had come up to me and given me NCP modules claiming they were great but in reality, they normally caused little problems but otherwise were unremarkable. This NCP module claims to be made in Singapore, and further digging seems to show that NCP was a brand of Hexon Technology Pte. Ltd. All the chips on the modules are remarked, but this module seems to have been installed January of 2001, with the chips labelled for Week 48 of 2000 and with their own NP33S168128K-7.5. The PCB appears to have scattered capacitors, and resistor packs as expected, with the pins not forming perfectly rounded rectangles – it has a small “pinch” near the top. The PCB appears to also cater for double-sided capacities.

A close inspection sees that this PCB is a slight variant of the Nova PCB with pretty much the same pin shaping and the position of various capacitors and inductors. Because of the commoditization of SDRAM, chips and PCBs became almost interchangeable provided their designs were compliant.

Generic 4-chip Module

20160117-1109-2131 20160117-1109-2132

Towards the real end of SDRAM, we got these monsters. Because SDRAM manufacturing had improved so much, manufacturers generally focused their efforts on higher density chips, making the lower density chips much harder to obtain. As a result, upgraders looking to cheap out on smaller capacity sticks had a choice to make.

Price lists for 128Mb modules would often have a cheaper price, followed by (4-chip) and a more expensive price, followed by (8-chip). The reason is that the 4-chip module is built using higher density parts which often provoked compatibility issues due to its organization.

This particular module has Fujimax F16M160128T-75 branded chips in a (likely) 16M x 8 bits x 2 banks configuration, which appears to be just another generic rebadge. The chips appear dated Week 41 of 2004, and this PCB has no provision for ECC or double-sided configurations, nor contains any resistor packs on the signal lines.

ECC Modules

Because of the number of modules, we jump forward a bit to some large ECC modules. Unlike older SIMMs where memory protection was provided by parity, under SDRAM, a more sophisticated ECC algorithm is used which is capable of repairing single bit errors and reporting two-bit errors. This comes with the same overhead – i.e. parity has 9 bits for each 8 bits of data, and ECC has 72 bits for each 64 bits of data, namely a one in nine redundancy.

ECC modules were generally more rare and difficult to encounter as they were supported only by servers and high-end workstations, of which I had none. But luckily, one day, I managed to salvage a set of 512Mb ECC SDRAM modules from a server – keep in mind, these three following types of modules were found in the one machine, which alludes to the fact that ECC SDRAM modules may have also been hard to get, resulting in this mix.

Legend L6472P33-724ISC3C Variant 1

20160117-1111-2141 20160117-1111-2142

We can see the PCB is full size with quite a bit of space and a cut-out to allow for easier side-clip access. The first module is based around Infineon ICs, specifically HYB39S256800CT-7.5, with a 8Mbit x 8 bits x 4 banks organization, dated Week 50 of 2001. The PCB is fully loaded with 18 chips in total.

Interestingly, Infineon were the semiconductor operations of Seimens (I had some of their SIMMs prior), and were later spun out to Qimonda which went bankrupt in 2009. At a time, Qimonda also had strategic alliances with Nanya Technology under the name of Inotera.

Legend L6472P33-724ISC3C Variant 2

20160117-1112-2147 20160117-1112-2148

The second variant came from the same system and has the same coding, but instead uses Hynix HY57V56820BT-H chips in its construction with the same organization. The chips were dated Week 52 of 2002. The vendor label is missing, but it’s likely it may have been peeled off or fell off at some point.

Legend L6472P33-724HSC3B

20160117-1111-2143 20160117-1111-2144

This one does have vendor labelling and is labelled Week 28 of 2002, and thus is a little bit earlier. The change of part number seems to be an artificial distinction – maybe there was a difference in SPD data for compatibility reasons, but the chips are essentially the same, as is the PCB. According to the label, I would assume the system was delivered on 27th February 2003, and thus all modules were probably installed at the time of delivery.

More to come …

That’s all for the first part. Join us in the next part to see my collection of 256Mb and 512Mb modules.

About lui_gough

I'm a bit of a nut for electronics, computing, photography, radio, satellite and other technical hobbies. Click for more about me!
This entry was posted in Computing, Tech Flashback and tagged , , , , , , . Bookmark the permalink.

4 Responses to Tech Flashback: My Collection of SDRAM – 128Mb & ECC Modules

  1. yuhong says:

    “This particular module has Fujimax F16M160128T-75 branded chips in a (likely) 16M x 8 bits x 2 banks configuration”
    The chips are 16M x 16 most likely I think. They were also common on SO-DIMMs BTW.

    • lui_gough says:

      Yeah, 16Mx16 is the same as 16Mx8 bitsx2 banks – i.e. the 8 bits x 2 banks is where the 16 comes from if I am not mistaken.

      – Gough

      • yuhong says:

        That is wrong. I believe that 64Mbit and higher chips have 4 banks, for one thing.

        • lui_gough says:

          Hmm. I’m not so sure then. Actually, I have made a mistake – the first thing is that SDRAM modules must have a 64-bit width, so for the 4-chip module, each chip has a width of 16-bits guaranteed. Of course, the whole module needs to sum to 128MiB/1024MBits. As a result, each chip needs to store 256Mbits. As a result, the only possibilities are: 16Mbits x 16 bits x 1 bank, 8Mbits x 16 bits x 2 banks, or 4Mbits x 16 bits x 4 banks to make it fit the bus width and total module capacity numbers.

          I looked around, and indeed, looking at Micron’s 256Mbit SDRAM range, the only 16-bit wide RAM has a geometry of 4Mbits x 16 bits x 4 banks. There is an 16 x 4 x 4 or 8x8x4 option, but neither will fulfill the 64-bit bus width required of the module –

          So indeed, I was wrong because I didn’t take into account the module bus width – so it’s 16M x 16 because it’s 4Mbits x 4 banks.

          Thanks for the information.

          – Gough

Error: Comment is Missing!