Samsung Begins Mass Production of HBM2 DRAM – 4 GB HBM2 For HPC, 8 GB HBM2 Production Commences This Year

Author Photo
Jan 19, 2016

Last year, Samsung announced that it will be commencing mass production of their HBM DRAMs in early 2016. Today,  along with advanced graphics, network systems and enterprise servers. Today, Samsung has officially commenced mass production of their first 4 GB HBM2 DRAMs which are going to target the HPC platforms for high-end graphics rendering and deep-learning capabilities.

gskill-tridentz-16-gb-3600_7-customRelatedG.Skill Unleashes Blistering Fast Trident Z DDR4-4600 MHz Memory – Designed For Kaby Lake-X, X299 HEDT Systems

Samsung Commences Mass Production of HBM2 DRAM – 4 GB HBM2 With 256 GB/s Bandwidth

The first product designed in accordance with the High Bandwidth Memory specifications is a 4 GB DRAM that is designed with the latest 20nm technology, offering better efficiency over previous generation DRAM chips. The HBM2 memory chips designed by Samsung will be integrated on next generation HPC systems and graphics cards that are tasked to run parallel computing, graphics rendering and machine learning. The production of the latest HBM2 DRAMs also marks a milestone for Samsung in the TSV (Through Silicon Vias) technology that is a crucial component of the HBM DRAM design.

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 big improvement in data transmission performance compared to typical wire-bonding based packages. via Samsung

The new HBM2 memory makes use of beefier 8Gb (Gigabit) dies which are connected and stacked vertically through the TSVs. Each package of these dies is able to deliver speeds of 256 GB/s (Gbps bandwidth) which is twice the bandwidth of current generation HBM1 memory that provides a 128 GB/s bandwidth and over 7x increase compared to a 4 Gb GDDR5 DRAM chip (36 Gbps). The 4 GB HBM2 package can be incorporated in several stacks, offering higher memory space at better efficiency than current generation DRAM solutions. Having either two of these packages on an interposer would get 8 GB of VRAM along with 512 GB/s of total bandwidth and four packages would mean 16 GB of VRAM and 1 TB/s bandwidth.

In previous slides, Samsung demonstrated that users can achieve a list of various designs with 4 GB HBM2 DRAM that can include chips with up to 24 GB HBM2 DRAM clocking in at 1.5 TB/s bandwidth. As for even higher memory demands, Samsung did acknowledge that they are well on their way to producing 8 GB HBM2 DRAM packages within this year to allow space savings of more than 95 percent, compared to using GDDR5 DRAM, offering more optimal solutions for compact devices that require high-level graphics capabilities in the consumer and enterprise sectors.

“By mass producing next-generation HBM2 DRAM, we can contribute much more to the rapid adoption of next-generation HPC systems by global IT companies,” said Sewon Chun, senior vice president, Memory Marketing, Samsung Electronics. “Also, in using our 3D memory technology here, we can more proactively cope with the multifaceted needs of global IT, while at the same time strengthening the foundation for future growth of the DRAM market.” via Samsung

nvidia-volta-3RelatedNVIDIA Working To Embed Multiple GPUs and Multiple Layers of Stacked DRAM on A Single MCM Package – Up To 256 SMs and Multiple Terabytes of Bandwidth

Image Credits/Courtesy: Computerbase

From the previous slides, we know that moving forward, Samsung will have a range of solutions packed in either 2-Hi, 4-Hi or 8-Hi stacks. The entry level models include the 2-Hi DRAM model that will be integrated on mainstream 2 GB HBM graphics cards (256 GB/s), performance based graphics with 4 GB HBM (512 GB/s). The Enthusiast graphics cards will ship with 4-Hi DRAM with 2 HBM stacks that will allow 8 GB VRAM (512 GB/s) and finally, 4 HBM Stacks with 16 GB VRAM models (1 TB/s).

On the HPC front, there are a wide array of high bandwidth and dense memory designs that include 4-Hi DRAMs with 4 HBM stacks that feature 32 GB VRAM (1 TB/s) and the bulky, 8-Hi DRAMs configured in 6 HBM stacks with 24 GB and 48 GB VRAM, both models featuring 1.5 TB/s bandwidth. There are also some network oriented HBM SKUs which are planned for launch in 2017 with 8-Hi DRAM Stacks configured in 1-2 HBM chips. In 2018, Samsung wants to focus on increase market growth by entering new applications to incorporate their HBM designs.

HBM2 is going to be the next generation DRAM standard to power upcoming iteration of graphics cards from AMD and NVIDIA. Codenamed Polaris and Pascal, the flagship GPUs of these families will be using over 1 TB/s HBM2 chips with VRAM of up to 32 GB. For more information on HBM2, you can read the official specification sheet of HBM2 published by JEDEC.

GDDR5 vs HBM Comparison:

HBM2 Specifications Comparison:

I/O (Bus Interface)326410241024
Prefetch (I/O)81622
Maximum Bandwidth32GB/s
(8Gbps per pin)
64 GB/s
(16Gbps per pin)
(1Gbps per pin)
(2Gbps per pin)
tCCD2ns (=4tCK)2ns (=4tCK)2ns (=1tCK)2ns (=1tCK)
VPPInternal VPPInternal VPPExternal VPPExternal VPP
VDD1.5V, 1.35V1.35V1.2V1.2V
Command InputSingle CommandSingle CommandDual CommandDual Command