This yet unnamed memory technology is able to transfer between 4.8 and 6.4Gbps per pin, about the same as XDR2 and GDDR5. Since XDR2 is differential and GDDR5 has some differential pins, this new methodology could be net more efficient per pin than the older two. In and of itself, that is nothing new, and it will vary by end user implementation.
The power levels are nothing new either, they are roughly in line with DDR3 per Gbps transferred, so what’s the point? Getting those speeds at that power level was previously not possible, you could do GDDR5 and burn watts, or DDR3 with a fraction of the bandwidth. This new tech can do both, and as you can see below, the signal is fairly clean while doing so.
Rambus’ new baby in action
Rambus gets there by using a 200mv swing on the signals, pretty low all things said. The fact that they can get it to work correctly and reliably is the proverbial magic, that is not something easy. How do they do it? That is the secret sauce, but they did it. Not just that, they did it with standard DIMM sizes, from large on the test rig to low profile and SO-DIMMs on the far left.
To make things more impressive, the DIMM that is running is using a standard off the shelf DDR3 socket. Also note that it uses a socket, GDDR5 has to be soldered down to reach those speeds. A socket, not to mention a dirt standard DDR3 DIMM socket, is quite the engineering feat. Don’t underestimate the problems associated with doing this over that many electrical transitions, it isn’t easy.
For the first time in a long time, Rambus has a high speed memory technology that is not differential, not high powered, and can use available connectors. While there are other forms of high speed memory technology imminent, DDR4 is the short term future, this could be better than anything proposed. It will be interesting to see if anyone picks it up, and for what.S|A
Latest posts by Charlie Demerjian (see all)
- AMD puts massive SSDs on GPUs and calls it SSG - Jul 25, 2016
- AMD revamps workstation lines and adds three cards - Jul 25, 2016
- AMD unveils Loom and ProRender software - Jul 25, 2016
- Emulex puts NVMe over FC with NVMf - Jul 20, 2016
- ARM needed Softbank, Softbank wanted ARM - Jul 19, 2016