SK Hynix Reveals HBM3 Memory Details – Capable of 665 GB/s Bandwidth
11 June 2021
AMD FURY NANO FURY X GDDR GDDR6 HBM HBM2 HBM2E HBM3 High Bandwidth Memory MEMORY R9 FURY Radeon Radeon VII RX VEGA SK HYNIX VEGA 56 VEGA 64

High Bandwidth Memory, developed in conjunction between SK Hynix and AMD has served the purpose of providing insanely fast, stackable DRAM modules close to the GPU core of which both are placed on an interposer. HBM memory provides vastly superior peak bandwidth and efficiency over traditional GDDR memory.

HBM started from humble beginnings first showcased on the AMD Radeon Fury X, though the overall memory capacity just wasn’t up to snuff (4GB) the Fury X showcased the benefits of utilizing HBM memory as it trounced the GeForce GTX 980Ti when it came to memory bandwidth at 512 GB/s effectively.

The proof of concept of utilizing this stackable DRAM has since been expanded upon through continued development with HBM2 memory being accessible in consumer grade products such as the Radeon RX Vega, with HBM2 providing a leap in capacity per virtual stack compared to first generation memory by up-to 8 times.

However continuing onwards from HBM2 memory was the introduction of HBM2E memory which further increased the maximum capacity and bandwidth per chip. While it may have been in mass production for over a year at this point HBM2E never really saw the light of day in any sort of consumer hardware which is actually quite depressing as even AMD have gone back to utilizing GDDR(6) memory since the introduction of the RDNA architecture, though this is mainly due to the fact that it no longer really relies upon massive peak bandwidths to get by thanks to the innovative Infinity Cache design in Radeon RX 6000 series graphics cards.

But that’s not to say that HBM memory is a thing of the past, more rather High Bandwidth Memory has never been stronger with SK Hynix announcing proposed specifications of next generation HBM3 memory.

Currently the HBM2E memory provided by SK Hynix provide a pretty staggering 460 GB/s per stack which is good for 3.6Gbps per pin and with the maximum capacity of HBM2E being set for 16 GB per chip.

However with SK Hynix’s HBM3 memory those figures will skyrocket to an incredible 665 GB/s (or greater) per stack, a mere 44.5% increase in bandwidth with an I/O speed of 5.2Gbps which is a 44.4% increase over HBM2E memory.

Though these figures are indeed subject to change as time continues forward but at the very least these are merely an illustration of the currently capabilities of SK Hynix’s HBM3 memory.

Don’t expect to see HBM3 memory in any sort of consumer based hardware anytime soon, however it might just make an appearance once AMD and NVIDIA move forward with MCM chiplet based GPU’s. Though do expect to see HBM3 memory in enterprise computing solutions once HBM3 is being readily mass produced of which we currently have no intended date for that from SK Hynix I’m afraid.

blog comments powered by Disqus