Samsung has developed the world's first 128 GB DDR4 registered memory modules for servers. From the press release:
Following Samsung's introduction of the world-first 3D TSV DDR4 DRAM (64GB) in 2014, the company's new TSV registered dual inline memory module (RDIMM) marks another breakthrough that opens the door for ultra-high capacity memory at the enterprise level. Samsung's new TSV DRAM module boasts the largest capacity and the highest energy efficiency of any DRAM modules today, while operating at high speed and demonstrating excellent reliability.
From The Register:
The Register is aware of servers with 96 DIMM slots, which means ... WOAH! ... 12.2 terabytes of RAM in a single server if you buy Samsung's new babies.
Samsung says these new DIMMS are special because "the chip dies are ground down to a few dozen micrometers, pierced with hundreds of fine holes and vertically connected by electrodes passing through the holes, allowing for a significant boost in signal transmission."
There's also "a special design through which the master chip of each 4GB package embeds the data buffer function to optimise module performance and power consumption."
Related Stories
Samsung has announced the mass production of dynamic random access memory (DRAM) packages using the second generation High Bandwidth Memory (HBM2) interface.
AMD was the first and only company to introduce products using HBM1. AMD's Radeon R9 Fury X GPUs featured 4 gigabytes of HBM1 using four 1 GB packages. Both AMD and Nvidia will introduce GPUs equipped with HBM2 memory this year. Samsung's first HBM2 packages will contain 4 GB of memory each, and the press release states that Samsung intends to manufacture 8 GB HBM2 packages within the year. GPUs could include 8 GB of HBM2 using half of the die space used by AMD's Fury X, or just one-quarter of the die space if 8 GB HBM2 packages are used next year. Correction: HBM2 packages may be slightly physically larger than HBM1 packages. For example, SK Hynix will produce a 7.75 mm × 11.87 mm (91.99 mm2) HBM2 package, compared to 5.48 mm × 7.29 mm (39.94 mm2) HBM1 packages.
The 4GB HBM2 package is created by stacking a buffer die at the bottom and four 8-gigabit (Gb) core dies on top. These are then vertically interconnected by TSV holes and microbumps. A single 8Gb HBM2 die contains over 5,000 TSV holes, which is more than 36 times that of a 8Gb TSV DDR4 die, offering a dramatic improvement in data transmission performance compared to typical wire-bonding based packages.
Samsung's new DRAM package features 256GBps of bandwidth, which is double that of a HBM1 DRAM package. This is equivalent to a more than seven-fold increase over the 36GBps bandwidth of a 4Gb GDDR5 DRAM chip, which has the fastest data speed per pin (9Gbps) among currently manufactured DRAM chips. Samsung's 4GB HBM2 also enables enhanced power efficiency by doubling the bandwidth per watt over a 4Gb-GDDR5-based solution, and embeds ECC (error-correcting code) functionality to offer high reliability.
TSV refers to through-silicon via, a vertical electrical connection used to build 3D chip packages such as High Bandwidth Memory.
Update: HBM2 has been formalized in JEDEC's JESD235A standard, and Anandtech has an article with additional technical details.
Previously:
AMD Teases x86 Improvements, High Bandwidth Memory GPUs
AMD Shares More Details on High Bandwidth Memory
Samsung Mass Produces 128 GB DDR4 Server Memory
Samsung Electronics has announced the production of "10nm-class" 8 gigabit DRAM chips that will be used in DDR4 modules with capacities ranging from 4 GB to 128 GB. "10nm-class" is an industry term that refers to an unspecified process somewhere between 10 nanometers and 19 nanometers.
In November, Samsung announced the production of 128 GB DDR4 registered dual inline memory modules (RDIMMs) using through silicon via (TSV) stacked dies with four 8 gigabit chips per package. Those modules used 20nm process DRAM and achieved a 2,400 Mbps data rate. The new 10nm-class memory will support a 3,200 Mbps data rate.
Samsung has demonstrated 64 GB RDIMMs using 16 Gb DDR4 memory chips, and plans to make 128 GB and 256 GB modules later this year:
Samsung is demonstrating its 64 GB DDR4 memory module based on 16 Gb chips this week at the OCP U.S. Summit. The 64 GB RDIMM that the company is showcasing is designed for mainstream servers, but ultimately the design will lend itself to build 128 GB and 256 GB memory modules for high-performance servers, the company said.
Samsung's monolithic 16 Gb DDR4 DRAM chips are rated for DDR4-2666 at the industry-standard 1.2 V. The chips are produced using an advanced manufacturing technology, but Samsung does not disclose details at the moment (it is logical to expect Samsung to use its '10-nm-class' tech though). The only thing we do know is that the fabrication process and monolithic die enable 20% lower power consumption of the demonstrated 64 GB RDIMM when compared to a module of the same capacity based on 8 Gb DDR4 chips.
In addition to the new dual-rank 64 GB RDIMM module, Samsung is set to develop quad-ranked 128 GB RDIMMs and octal-ranked 256 GB LRDIMMs. Today's servers running AMD's EPYC or Intel's Xeon Scalable M-suffixed processors feature 12 or 16 memory slots - if the processors were capable of fitting all 256 GB modules, this could lead up to 4 TB per socket. This should be a massive advantage for applications like in-memory databases, virtual desktop infrastructure, and so on.
16 Gb chips may also end up being used in 32 GB memory modules for desktop users.
DDR4 RDIMM and LRDIMM Performance Comparison
Also at Samsung.
Related: Samsung Mass Produces 128 GB DDR4 Server Memory
Samsung's plans to make 256 GB memory modules using 16 Gb chips are moving forward:
Samsung this week demonstrated its first 256 GB memory module for upcoming servers. The new Registered DIMM (RDIMM) is based on Samsung's 16 Gb DDR4 memory devices introduced earlier this year and takes advantage of the company's 3DS (three-dimensional stacking) packaging. The new module will offer higher performance and lower power consumption than two 128 GB LRDIMMs used today.
Samsung's 256 GB DDR4 Registered DIMM with ECC carries 36 memory packages featuring 8 GB (64 Gbit) of capacity each, along with IDT's 4RCD0229K register chip (to buffer address and command signals and increase the number of ranks supported by a memory channel). The packages are based on four single-die 16 Gb components that are interconnected using through-silicon vias (TSVs). Architecturally, the 256 GB module is octal ranked as it features two physical ranks and four logical ranks.
1 TB can't be too far behind.
Previously: Samsung Mass Produces 128 GB DDR4 Server Memory
Samsung Shows Off New 64 GB Server Memory Modules Using 16 Gb Chips, Promises 128-256 GB This Year
Samsung Unveils 32 GB Laptop DDR4 Modules
(Score: 2, Informative) by Anonymous Coward on Friday November 27 2015, @03:41AM
DDR4 SDRAM RDIMMs using TSVs (Through silicon vias) with stacked dies (4 deep). Built on the 20nm process, 8 gb per die (4GB per package). This version is only 2,400 megabits per second (They mean per wire: that's the effective bus speed (double the clock speed), which isn't great for DDR4) but faster (and higher power) versions are planned.
I think its worth noting that not all computers can simply have all their ram slots filled with massive ram: often there is a limit on the amount of supported ram.
The summery is phrased like we both don't know what through silicon vias are, or what registered/buffered memory (R in RDIMM) means: If you are going to provide a dumbed down explanation, please also provide the technical term for a nice combo of clarity+education.
(Score: 1, Informative) by Anonymous Coward on Friday November 27 2015, @06:12AM
I think its worth noting that not all computers can simply have all their ram slots filled with massive ram: often there is a limit on the amount of supported ram.
FWIW, intel's Xeon E7 arch has 46-bit physical addressing [intel.com] which works out to 64TB of ram.
(Score: 4, Interesting) by ledow on Friday November 27 2015, @09:55AM
Not being funny, but I'm a tech guy and... I don't really care.
The TSV stuff? Makes no difference to the way the chip operates as far as I'm concerned. I know what they are, but that's it.
Stacked dies? Cool science and everything but, again, why would I care?
20nm process? I'm sure the number means something to some people. I know what it means, represents, etc. but in a product it's just a marketing statistic.
2,400 megabits per second. That means something. That's a number that may affect me. With caveats as you've pointed out.
But the critical point is your middle one - can I just slap this in a computer and have it work? Probably not. And in terms of server things, it may even be years before someone like Lenovo etc. certify this kind of tech to go into my Blade servers.
Honestly, although the details sound cool, they are literally just trivia. Samsung could well be making them up, is anyone checking or even able to check?
But it's a 128Gb RDIMM at a certain speed. That's all most people, even most people building servers for a living, ever need to know.
As such, most techies probably find the explanations handy for the first time they encounter the terms, but its' really nothing more than a trivia factlette to know, or care, or bother to find out.
It's like processors and processor sockets nowadays. At one time I knew them all and could tell people all kinds of useless information. Now that I buy computers in the hundreds, I can't remember the last time I *ever* had to research what socket a motherboard had, or what processors were compatible with that. I honestly couldn't care, and my suppliers sort it out for me when I tell them I need 8-cores with this much RAM and these ports on the back. At worst, I would find out the "code number" of that socket on a motherboard I like, google the processors that fit in that, and when I find the combination I like, I would buy it. What the name, technology, size or whatever other statistics of that socket is? Who cares? I have little clue what the Intel numbering system is at all nowadays, it has almost zero correlation to the features or speeds of the chips, and motherboard sockets are even worse.
Honestly, who here really buys server RAM on the basis of anything other than "128Gb, please, and make it the fast stuff, and buy the proper IBM/whatever stuff so it doesn't void my warranty"?
(Score: 1, Interesting) by Anonymous Coward on Friday November 27 2015, @01:14PM
Yep, your long rant shows that you really, really, you really mean it, really, don't care.
Bug off.
(Score: 2) by takyon on Saturday November 28 2015, @05:24AM
These modules are 128 gigabytes, not gigabits. 128 Gb = 16 GB.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Interesting) by darkfeline on Friday November 27 2015, @08:40AM
Wonder what the power cost will be simply refreshing those 12.2 terabytes of RAM? That's a lot of energy gone simply keeping the memory active.
Join the SDF Public Access UNIX System today!
(Score: 2) by Gravis on Friday November 27 2015, @01:10PM
less power than it takes to power the screen on your cellphone.
(Score: 3, Interesting) by rob_on_earth on Friday November 27 2015, @10:30AM
So when can I buy a motherboard for a desktop machine that will take these?
And before you ask I love playing with Blender 3D
In fact what is the most memory you can get on self build desktop at the moment?
I remember looking a good few years ago and all the motherboards had weird restrictions either on the total size per module or the number of modules.
(Score: 3, Informative) by gnuman on Friday November 27 2015, @05:11PM
You know, these are trivial to investigate. You can easily get 64GB on a desktop these days. 8GBx8.
AMD: 64GB
http://www.newegg.com/Product/Product.aspx?Item=N82E16813157479 [newegg.com]
Intel: 128GB
http://www.newegg.com/Product/Product.aspx?Item=N82E16813157543 [newegg.com]
And these are cheap options too. And what is the problem with Blender? Do you honestly need more than this for your typical work?
(Score: 3, Interesting) by twistedcubic on Friday November 27 2015, @08:12PM
Doesn't seem like those support ECC. Who would use 64GB+ of non-ECC memory?
(Score: 0) by Anonymous Coward on Friday November 27 2015, @01:46PM
So, any empirical data as to how they handle heat dissipation?
(Score: 2) by takyon on Saturday November 28 2015, @05:22AM
I don't expect empirical data until they reach market. You want independent testing anyway. There has been plenty of research on how to improve TSV thermal management, and you can expect that to continue.
http://www.ansys.com/staticassets/ANSYS/staticassets/resourcelibrary/article/AA-V4-I2-Avoiding-Excess-Heat-in-TSV-Based-3D-IC-Designs.pdf [ansys.com]
http://gradworks.umi.com/35/25/3525579.html [umi.com]
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]