SK Hynix was one of the main developers with the original HBM memory back in 2014, and the company wants to stay ahead of the curve with this premium type of DRAM. Embedded within the company’s 1 billion fab process qualification memo on Tuesday, the manufacturer said it is working on next-generation HBM3E memory that will enable speeds of up to 8 Gbps/pin and will be available from 2018. revealed for the first time. 2024.
The latest HBM3 memories from SK Hynix and other vendors support data transfer speeds up to 6.4Gbps/pin, so HBM3E with 8Gbps/pin transfer speeds can provide a bandwidth advantage over existing memory devices. provides a moderate 25% advantage in
To put this in context, a single HBM stack using a 1024-bit wide memory bus yields a known good stack die (KGSD) bandwidth for HBM3E of about 1 TB/s, and 819.2 GB/s for the following: seconds to maximum. HBM3 today. Using modern HPC-class processors that employ 6 (or more) stacks, this translates to several TB/s of bandwidth for high-end processors.
According to the company’s memo, SK Hynix plans to start sampling the HBM3E memory next month and start mass production in 2024. The memory maker didn’t reveal much about the details of his HBM3E (in fact, this is the first public release since its specs aren’t mentioned at all), but it’s likely that these devices will be compatible with his existing HBM3 controller and physical interface. I’m not sure if it’s drop-in compatible.
| HBM memory comparison | ||||||
| HBM3E | HBM3 | HBM2E | HBM2 | |||
| Maximum capacity | ? | 24GB | 16 GB | 8GB | ||
| Maximum bandwidth per pin | 8Gb/s | 6.4Gb/s | 3.6Gb/s | 2.0Gb/s | ||
| Number of DRAM ICs per stack | ? | 12 | 8 | 8 | ||
| effective bus width | 1024 bits | |||||
| Voltage | ? | ? | 1.2V | 1.2V | ||
| Bandwidth per stack | 1TB/s | 819.2GB/s | 460.8GB/s | 256GB/s | ||
Assuming SK Hynix’s HBM3E development is going according to plan, the company should have little trouble arranging customers for even faster memory. In particular, demand for GPUs used to build AI training and inference systems has skyrocketed, prompting NVIDIA and other processor vendors to develop the advanced technology needed to produce faster processors during this boom in the industry. We are willing to pay a premium for extra memory.
SK Hynix plans to use the technology to produce HBM3E memory. 1b Nanometer manufacturing technology (5th generation 10nm class node). It is currently being used in the production of his DDR5-6400 memory chips that will be validated for Intel’s next-generation Xeon Scalable platform. In addition, this manufacturing technology is also used to produce his LPDDR5T memory chips, which combine high performance and low power consumption.
