Rambus DRDRAM (or RDRAM as it was sometimes known) was a proprietary memory technology championed by Rambus and Intel as a (potential) successor to SDRAM. As memory bandwidth requirements rose in the early Pentium 4 era, there was a need to ensure the CPU continued to be well fed with data to keep it occupied. Rambus technology memory offered a solution to the problem by running the bus with data on both rising and falling edges, a technology we term “double data rate” or DDR.
Interestingly, this happened around the same time that JEDEC who produces the majority of memory-related standards was trying to standardize a higher performance DDR-based SDRAM, and this led to a lot of accusations in the industry of deliberate attempts to patent technologies and extort unreasonable licensing fees, and attempts to circumvent patents to avoid paying licensing fees. It was a very messy time.
But at the time, Rambus’ success almost looked to be assured. Intel had decided solely to support Rambus in several chipsets for their Pentium 4 CPUs. With such a might behind the Rambus technology, and a decision to not support SDRAM (DDR or otherwise) at all, Intel subsidised Rambus quite heavily to try and make it the market standard.
For a brief moment, Rambus memory was the highest performing option for consumer systems, and server/workstation builds were well advised to use it. This was despite the high cost, high heat, and oddities of the system.
For one, most of the modules looked like this, with metal heatspreaders riveted onto the PCB. These were non-removable and were absolutely necessary unlike some of the heatspreaders on today’s modules. The reason was that Rambus operated like a bus and each of the chips inside were complete with their own memory bus logic which made them more expensive and power consuming.
Underneath the heat spreaders (on a different set of RIMMs) are shiny chips which are basically the underside of the silicon die poking up. In fact, that’s what the logo for this site is made out of. As each chip talked to its neighbours in a bus formation, this increased latencies and put limitations on how modules could be combined. Empty slots, for example, often had to be filled with continuity RIMMs, or CRIMMs.
This particular set of CRIMMs comes from a Dell server, and basically connects one set of incoming pins to the other set of pins on the other side. The actual circuit pattern is hidden because it’s a 6-layer board with planes on both sides to ensure signal integrity and prevent signals radiating out.
Unfortunately (or, rather fortunately), soon after, Rambus memory was not competitive on a latency, cost, heat and performance basis compared to DDR and was rarely used in marketed platforms with the rare exception of a gaming console. It was completely outpaced by the JEDEC-backed DDR standard at higher clock frequencies which had less of the annoyances.
Rambus (the company) itself became synonymous with being a nasty patent troll constantly dragging DRAM manufacturers through litigation processes and attempts to settle on alleged patent infringements. I suppose initially, it is easy to feel sorry for Rambus, but their later behaviour seemed just as provocative as that of the defendants. I suppose in all, JEDEC’s victory with DDR (and successive generations) becoming the standard memory technology has benefited the consumer in the long run by keeping proprietary memory technologies at bay, and enabling an industry standard where multiple manufacturers are free to compete directly, and device designers can have comfort in knowing that all the available parts are (for most intents) directly interchangeable. This is something to remember, especially as we are slowly creeping back to a world of proprietary standards and vendor lock-in/DRM/etc.
Bonus: Interchangeable VRM Module
As it turns out, the Dell branded workstation/server that the RIMMs were pulled out of also had a curious feature of an “interchangeable” VRM module for the CPU. I suppose this would have allowed easy replacement of a VRM if it failed in the field without replacing the whole motherboard, or upgrading of a VRM if a later CPU required more sophisticated voltage planes (depending on how flexible the design was).
This particular card has an edge connector that has pin-spacing identical to an ISA card, but with different length and keying. The card itself powered a Pentium 4 1.4Ghz Socket 423 CPU, and appears only to have two of three possible phases populated. The main VRM controller, MOSFET driver and MOSFETs are an Intersil solution.
The inductors are fairly low profile and have a cemented ferrite cover for better efficiency. It reminds us that CPUs of the early Pentium 4 days were relatively limited in power consumption – the CPU claiming a gentle 54.7W TDP.
As a result, even the VRM doesn’t need a particularly beefy heatsink, with this one mounted on the rear, relying on the vias on the board to carry heat through to the rear (i.e. a not very efficient way of cooling). The board appears to be a product of Artesyn Technologies who specialize in embedded power solutions.