Samsung Electronics Supplying Twelve Layer HBM4E Memory Samples To Respond To Large Scale Processing Demands Of AI Data Centers
Samsung Electronics has started supplying first samples of their new 12 layer HBM4E memory to global corporations. Official company press releases state the new memory architecture is built upon the production and deployment of the earlier HBM4 version to respond to the increasingly large scale processing demands of artificial intelligence data centers and hyperscale computing facilities.
The new HBM4E hardware will run at 14 Gbps at constant pin speed but can go to 16 Gbps at maximum workload intensity. This results in over a 20% improvement from HBM4 performance and each stack is capable of a 3.6 TB/s bandwidth to speed up the processing of large language models. This 12 layer version holds 48 GB of data while having over a 30% density increase from older generations of the product. The company is expected to continue the line to offer an 8 layer 32GB configuration and 16 layer 64GB configurations.
To produce this new memory, Samsung is using their newest 6th generation 10nm class DRAM process with a 4nm logic base die produced by Samsung Foundry. This internal production process ensures a stable and high yieldrate. Energy efficiency gains have resulted in over 16% less energy consumption and over a 14% decrease in thermal resistance than older versions of the product, allowing for high density workloads to run at lower temperatures and cooler.
"Following the successful mass production of HBM4, Samsung has once again demonstrated its distinct technological edge with HBM4E. Through our advanced manufacturing capabilities and preemptive infrastructure investments, we will continue to drive the growth of the global AI memory market."
Sang Joon Hwang, Executive Vice President and Head of Memory Development at Samsung Electronics
System in package evaluations of the HBM4 architecture during previous testing showed speeds up to 11.7 Gbps, so utilizing the identical core and base die that was verified during the HBM4 production cycle should lead to swift mass production and readily meet demand for next generation computational hardware.
