Korean SK Hynix announced its new generation of HBM3 RAM. Designed to deliver ever higher transfer speeds, while benefiting from seriously refined energy efficiency, these modules will serve as video memory for future high-performance graphics cards.
Take over from the current HBM2E memory by improving just about everything. This is the reason to live from the HBM3 memory that SK Hynix presented.
Speed and power
This new generation of DRAM modules promises stunning transfer speeds, which can go up to 5.2 Gbps in input and output. An increase of 44% compared to HBM2E memory. Same dynamic for the bandwidth that will be able to rise to a maximum of 665 GB/s of transferred data against 460 GB/s currently. Here too we are faced with 44% improvement in WCCFTech rating, which offers us an interesting scale-up.
By doing the calculation, a single HBM3 module should be able to offer more bandwidth than the total proposed by the Radeon RX 6900XT and GeForce RTX 3070 Ti combined. Four modules of this new DRAM will also be sufficient to develop 2 TB/s of bandwidth while the six VRAM modules of the NVIDIA A100 (dedicated to supercomputers) are limited to 3 TB/s.
Planned launch next year
But speed is not everything and SK Hynx knows it. The brand promises an improvement of the dissipation technologies introduced with its HBM2E memory. The result: 14 degrees less on the modules and an improved heat dissipation of 36%, promises the South Korean group.
The maximum capacity of these new DRAM modules, however, is not expected to change significantly compared to HBM2E memory—at least for the first generation expected next year. Initially, we will find dies of 16 Gb for a capacity of 16 GB. However, the density of the modules may increase when the HBM3 specifications are finalized by the Joint Electron Device Engineering Council (JEDEC). It is therefore not excluded that this capacity will be revised upwards in the medium term.
What about the first graphics cards eligible for this new DRAM HBM3 memory? We have to look for future GPUs for servers and supercomputers from AMD, NVIDIA and Intel. WCCFTech cites the AMD Instinct, NVIDIA Hopper and Intel Xe-HPC designs, which will be among the first to be served.