Samsung, Server Memory, and Being Green

Microsoft’s Munich-based Technology Center and Samsung Electronics recently “announced availability of an optimized platform utilizing Samsung’s advanced Green Memory in Microsoft’s virtualized data center.”

Although Samsung’s announcement is important, one could read through the release and make some assumptions about the memory market that might not be entirely accurate. So let’s dig a little deeper and make sure the reasons why it is significant is apparent.

Samsung states that the 30nm “green” memory tested by Microsoft yields higher performance and lower power utilization in demanding Hyper-V environments, such as the one described in Microsoft’s white paper. But why is this? Does Samsung have some secret green sauce, or could any memory achieve the same thing. The answer is a bit of both.

Let’s bullet point what this release is really saying:

• Microsoft tested 30 nanometer (nm) DDR3 modules running 2 Gbit (Gb) components at 1333MHz, although the parts can reach up to 1866 MHz.
• Samsung’s current 30nm parts operate in accordance with the 1.35V JEDEC specification.
• The near future use of 4Gb rather than 2Gb chips means fewer chips on a given capacity memory module, which in turn contributes to power savings.
• Samsung’s 20nm silicon will draw less power than a similar product manufactured on a 30+ or 40+ nm fabrication node.

These are excellent points, and Samsung deserves praise for them. But Elpida recently herald its own 25nm 4Gb DDR3 rated at up to 1888MHz, which Elpida noted would deliver 25% to 30% lower operating current and 30% to 50% lower standby current than prior 30nm parts. Intel and Micron also announced 20nm NAND fabrications that were already sampling last April.

So there is plenty of competition for Samsung’s advances, and the South Korean giant may at times be announcing that it has caught up rather than forged the future. In a way, improved fabrication and going green are inextricably linked.

More broadly, we should ask what “green” memory might mean in the absence of an industry definition. Today’s server memory market is defined by JEDEC specifications, which defines DDR3L (for “low voltage”) as drawing 1.35V rather than the conventional 1.5V defined for DDR3, and pretty much every memory vendor has DDR3L-compliant models.

Then there’s the issue of fabrication node and transistor density. According to Mike Mohney, Technology Manager, Kingston, “If you have Chip A using a 20nm process and Chip B using a 30nm process, you will probably draw less power—maybe 45 or 50mA compared to 55 or 60mA. Multiply that times 1.35V and you’ll see you have less total power being consumed.”

With smaller transistors also comes higher storage density, hence 4Gb chips coming to replace 2Gb chips. A 4 gigabyte (GB) module requires fewer 4Gb chips than 2Gb chips, so again less power is consumed for the same memory capacity.

Samsung’s achievement is not only that it has done all these things but that it has done them first. Moreover, with Samsung verging on the 20nm node, the company has nearly caught up to Intel, signaling a dramatic shortening of the process gap that has traditionally separated memory from processor manufacturing. As with processors, the significance of this evolution inevitably boils down to faster speeds, higher capacities, and lower power consumption compared with the prior generation.

Businesses should be targeting low-power components in their systems, but they should understand what factors contribute to low power, not simply the fact that a given company tosses around the word “green” for marketing. This fact will take on added importance as JEDEC updates its server memory specs for 1600MHz and possibly beyond going into 2012.