Rambus aims for a terabyte of bandwidth

THIS WEDNESDAY, RAMBUS will be announcing its Terabyte Bandwidth Initiative (TBI), aimed at eventually hitting that magic TB figure. It all about bandwidth per device, not aggregate, and with the things it is showing soon, it can be achieved.

To put that in perspective, the PS3, a Rambus licensee, has about 50GB/s in bandwidth, so in three years, it aims to pump that up by 20x. To do this, it has three key technologies, 32x data rate, Flexlink C/A, and Fully Differential Memory Architecture.

The first one, 32x Data Rate is like DDR on steroids. DDR is dual data rate, or two bits per clock. Rambus has been pushing that envelope for a bit, and XDR memory uses 8x data rate, or 4x what DDR does. The new 32x Data Rate ups this to 16x DDR, per clock, and 4x that of XDR. With the same 500MHz base clock, you can get 16Gb/s per pin. It looks like this.

If you do the mathematics, a device with 16 DRAMs each at 16Gb/s with a width of 4B per device will get Rambus there. Think about this for graphics chips, and the like which today are only just breaking the 100GB/s barrier.

The next chunk of this is aimed at the command and control bits, pun intended. Flexlink C/A brings the same high speed links that the data bus has to the command/address link. Basically, instead of having a parallel bus for the command and address bits, you get a single high speed point to point link.

That means you only have to route one type of signal between the memory controller and the DRAM, theoretically simplifying things. You also need to route many fewer wires. If you need more granularity, you can just add more wires, something that was very hard the 'old way'. Rambus also pulled in some of the phase compensating technology from XDR, called "Flexphase", to work here.

FDMA is the last one, and it builds on the Flexlink C/A to get you a fully differential memory interface. DDR3 has only the clocks as differential signals, DQ and C/A are all single ended. XDR has differential DQ but C/A is still single ended. TBI makes it differential end to end.

On Wednesday, Rambus plans to demo this technology working on production targeted parts. It is not planning to use special materials or building techniques, just was is, or will soon be production processes.

If you remember, TBI is the Terabyte Bandwidth Initiative, the idea is to make things work in the real world, not an idea that would ever make it to volume. One thing is clear, bandwidth on this scale will be needed sooner rather than later, and if Rambus can demo it working in a few days, it stands a good chance of making it to devices in 3 or so years. Let's hope it gets there. µ