Thursday, December 19, 2024

Samsung’s 9.8 Gbps HBM3E: A Revolution in Memory!

- Advertisement -

Samsung Intros 9.8 Gbps HBM3E “Shinebolt”, 32 Gbps GDDR7, 7.5 Gbps LPDDR5x Cam2 Memory

During its Memory Tech Day 2023, Samsung formally unveiled its next-generation memory technologies, including HBM3E, GDDR7, LPDDR5x CAMM2, and more.

Samsung Makes a Major Investment in Next-Gen Memory Technologies, Including HBM3E, GDDR7, LPDDR5x, CAMM2, and More

- Advertisement -

For the newest in artificial intelligence, gaming, and data centre applications, Samsung HBM3E memory dubbed “Shine Bolt” and GDDR7 have already been reported on. These two events can be considered the Memory Tech Day 2023’s two main highlights, but Samsung is undoubtedly hosting a lot more events.

HBM Memory Specifications Comparison

DRAMHBM1HBM2HBM2EHBM3HBM3 GEN2HBMNEXT (HBM4)
I/O (Bus Interface)10241024102410241024-20481024-2048
Prefetch (I/O)222222
Maximum Bandwidth128 GB/s256 GB/s460.8 GB/s819.2 GB/s1.2 TB/s1.5 – 2.0 TB/s
DRAM ICs Per Stack488128-128-12
Maximum Capacity4 GB8 GB16 GB24 GB24 – 36 GB36-64 GB
tRC48ns45ns45nsTBATBATBA
tCCD2ns (=1tCK)2ns (=1tCK)2ns (=1tCK)TBATBATBA
VPPExternal VPPExternal VPPExternal VPPExternal VPPExternal VPPTBA
VDD1.2V1.2V1.2VTBATBATBA
Command InputDual CommandDual CommandDual CommandDual CommandDual CommandDual Command

“Shinebolt” Samsung HBM3E Memory For AI & Data Centres

Samsung today unveiled Shinebolt, its next-generation HBM3E DRAM, building on the company’s experience in commercialising the sector’s first HBM2 and developing the HBM market for high-performance computing (HPC) in 2016. Next-generation AI applications will be powered by Samsung’s Shinebolt, which will lower total cost of ownership (TCO) and hasten inference and training of AI models in the data centre.

The HBM3E can transfer data at rates greater than 1.2 terabytes per second (TBps) thanks to its remarkable 9.8 gigabits per second (Gbps) per pin performance. Samsung has optimised its non-conductive film (NCF) technology to remove gaps between chip layers and maximise thermal conductivity in order to enable higher-layer stacks and improve thermal characteristics. Shinebolt samples are presently being sent to clients, and Samsung’s 8H and 12H HBM3 devices are currently in mass manufacturing.

The business also intends to provide a unique turnkey service that integrates next-generation HBM, cutting-edge packaging technologies, and foundry options, leaning on its position as a whole semiconductor solutions supplier.

- Advertisement -

GDDR Graphics Memory Evolution:

GRAPHICS MEMORYGDDR5XGDDR6GDDR6XGDDR7
WorkloadGamingGaming / AIGaming / AIGaming / AI
Platform (Example)GeForce GTX 1080 TiGeForce RTX 2080 TiGeForce RTX 4090GeForce RTX 5090?
Number of Placements12121212?
Gb/s/pin11.414-1619-2432-36
GB/s/placement4556-6476-96128-144
GB/s/system547672-768912-11521536-1728
Configuration (Example)384 IO (12pcs x 32 IO package)384 IO (12pcs x 32 IO package)384 IO (12pcs x 32 IO package)384 IO (12pcs x 32 IO package)?
Frame Buffer of Typical System12GB12GB24 GB24 GB?
Average Device Power (pJ/bit)8.07.57.25TBD
Typical IO ChannelPCB (P2P SM)PCB (P2P SM)PCB (P2P SM)PCB (P2P SM)

Next-Generation Gaming Graphics With 32 Gbps & 32 Gb DRAM From Samsung

Other products shown at the event include the 32Gb DDR5 DRAM with the greatest capacity in the industry, the first 32Gbps GDDR7 in the market, and the petabyte-scale PBSSD, which significantly improves server application storage capabilities.

Samsung claims that GDDR7 memory will perform 40% better and use 20% less power than the 24 Gbps GDDR6 DRAM currently available, which has a 16 Gb die capacity. A 384-bit bus interface solution will be used to achieve up to 1.5 TB/s of bandwidth with the first devices, which will be rated at transfer speeds of up to 32 Gbps, a 33% improvement over GDDR6 memory.

The bandwidth that the 32 Gbps pin speeds would provide for various bus designs is as follows:

  • 2048 GB/s (512 bits) or 2.0 TB/s
  • 384-bit, 1.5 TB/s, 1536 GB/s
  • 1.3 TB/s, 320-bit, 1280 GB/s
  • 1024 GB/s (256 bits) and 1.0 TB/s
  • 768 GB/s at 192 bits
  • 512 GB/s at 128 bits

The business has also tested early prototypes with speeds of up to 36 Gbps, but we don’t think those will be ready in sufficient bulk quantities to fill the lineups for the next-generation gaming and AI GPUs.

Additionally, GDDR7 memory will be 20% more efficient, which is fantastic given how much power memory uses on high-end GPUs. According to reports, the Samsung GDDR7 DRAM will have technology that is especially suited for high-speed workloads, as well as a low-operating voltage alternative created for applications with careful power consumption, such laptops.

The new memory standard for thermals will make use of an epoxy moulding compound (EMC) with high thermal conductivity, which lowers thermal resistance by up to 70%. Back in August, it was revealed that Samsung was giving NVIDIA samples of their GDDR7 DRAM for an early assessment of its next gaming graphics cards.

Samsung LPDDR5x Reduces Mobile Designs and Next-Gen CAMM2 Modules

Today’s AI technologies are heading towards a hybrid paradigm that divides and distributes workload among cloud and edge devices in order to process data-intensive activities. As a result, Samsung unveiled a selection of memory solutions that support edge computing with high-performance, high-capacity, low-power, and tiny form factors.

The company also displayed its 9.6Gbps LPDDR5X DRAM, LLW DRAM specialised for on-device AI, next-generation Universal Flash Storage (UFS), and the high-capacity Quad-Level Cell (QLC) SSD BM9C1 for PCs, which is anticipated to be a true game changer in the next-generation PC and laptop DRAM market.

GPU Memory Technology Updates

GRAPHICS CARD NAMEMEMORY TECHNOLOGYMEMORY SPEEDMEMORY BUSMEMORY BANDWIDTHRELEASE
AMD Radeon R9 Fury XHBM11.0 Gbps4096-bit512 GB/s2015
NVIDIA GTX 1080GDDR5X10.0 Gbps256-bit320 GB/s2016
NVIDIA Tesla P100HBM21.4 Gbps4096-bit720 GB/s2016
NVIDIA Titan XpGDDR5X11.4 Gbps384-bit547 GB/s2017
AMD RX Vega 64HBM21.9 Gbps2048-bit483 GB/s2017
NVIDIA Titan VHBM21.7 Gbps3072-bit652 GB/s2017
NVIDIA Tesla V100HBM21.7 Gbps4096-bit901 GB/s2017
NVIDIA RTX 2080 TiGDDR614.0 Gbps384-bit672 GB/s2018
AMD Instinct MI100HBM22.4 Gbps4096-bit1229 GB/s2020
NVIDIA A100 80 GBHBM2e3.2 Gbps5120-bit2039 GB/s2020
NVIDIA RTX 3090GDDR6X19.5 Gbps384-bit936.2 GB/s2020
AMD Instinct MI200HBM2e3.2 Gbps8192-bit3200 GB/s2021
NVIDIA RTX 3090 TiGDDR6X21.0 Gbps384-bit1008 GB/s2022
NVIDIA H100 80 GBHBM3/E2.6 Gbps5120-bit1681 GB/s2022

News source

- Advertisement -
RELATED ARTICLES

Recent Posts

Popular Post

Govindhtech.com Would you like to receive notifications on latest updates? No Yes