Samsung bakes lower-voltage server memory
When every watt counts
Those hyperscale data centers that are trying to cram 10 pounds of servers into a 5-pound rack have gotten a little more help from Korean memory maker Samsung Electronics.
The leading edge, low-voltage DDR3 main memory runs at 1.35 volts and burns a bit less juice than normal 1.5 volt memory modules. And now, Samsung is pushing the power envelope down with its 1.25 volt DDR3 "green memory", which it was showing off last week at the Asia Pacific Datacenter Leadership Council in Singapore.
Like other DDR3 memory modules, the new green memory sticks, aimed specifically at ultra-low-power servers, are based on registered dual inline memory modules (RDIMMs). Like some of the existing 1.35 volt sticks, the 1.25 volt green memory uses what Samsung inscrutably calls its "30 nanometer class" fabrication processes, which means it can range from 30 to 39 nanometers. Other 1.35 volt modules are based on "40 nanometer class" processes, which means anywhere from 40 to 49 nanometers.
Up until now, a 16GB RDIMM based on the 30-nanometer-class processes and implemented using 4Gb memory chips running at 1.35 volts had the lowest power profile. But the 1.25 volt parts, which hum along at 1.33Mb/sec of bandwidth, consume only 3.7 watts – compared to 4.35 watts for the 1.35 parts of the same density. And 1.35 volt 16GB DDR3 parts – based on cheaper and older 2Gb memory chips etched with 40-nanometer-class processes – burn a wonking 9.25 watts. That means the new green memory from Samsung will consume 60 per cent less juice than these older parts.
And that sure does mount up, as Charlie Sheen once said about his exorbitant hooker bill. In the DDR2 days, memory capacity using 1Gb chips made in 60 nanometer processes and running at 1.8 volts would burn 102 watts to deliver 48GB of capacity using 4GB memory sticks. Now, for around 14.8 watts, you can cram the same 48GB into one-quarter the number of slots and run the memory considerably faster, too. It also means that within the same thermal budget, you could put memory capacity greater by nearly a factor of seven into a server. This is surely going to come in useful for fat servers running hypervisors and loading up dozens and dozens of virtual machines on their threads.
Samsung says it will wire the 1.25 volt memory chips into 4GB, 8GB, and 16GB after its OEM partners qualify early samples. Volume production is not expected until next year, and hopefully early enough to be concurrent with the next rev of server processors from Intel and Advanced Micro Devices. ®
"1.33Mb/sec of bandwidth" - my broadband is faster than that! I'll take the high power option
Memory Arms Race?
Looks like the race to low power between DRAM and SSD/Flash is getting exciting. Assuming that the early adopter premium goes away quickly, the cost per VM in a dual 16 or 12-core systems is going to drop a lot, and the ability to crank to 196GB of DRAM per U means 4x the VMs of a typical system today.
Flash for VM images makes sense, but we will have achieved storage IO starvation for operating traffic. I've been saying for some years that we will run into an IO wall. 2012 is when it happens!
Flash cache and other tricks are not enough to speed things up. We'll need some new transfer constructs and storage paradigms to move on.
Apple - No Chips 4 U
Most probably Samsung will address the evil empire in an appropriate an measured way.
And technology just keeps getting better!