SK Hynix Next-Gen HBM3 Memory Primed To Deliver A Blistering 665GB/s Of Bandwidth
High bandwidth memory (HBM) has always lived up to its name, it just has not been as widely adopted in mainstream graphics cards as GDDR memory chips. Maybe that will change when HBM3 arrives. Regardless, whatever products do end up getting the HBM3 treatment will see a big jump in bandwidth, according to some figures shared by South Korean memory make SK Hynix.
What's interesting is that SK Hynix is not even aggressively pushing the promise of HBM3 memory just yet. The bandwidth claim was mentioned as an aside, on a page dedicated to extolling the benefits of the company's HBM2E memory solution, which it is claims is the industry's fastest memory at 3.6Gbps and can process 460GB of data per second.
Good stuff, but HBM3 will blow HBM2E out of the water, in terms of bandwidth and I/O speed. By how much, exactly?
SK Hynix posted an info graphic (above) touting an increase to more than 665GB/s of bandwidth for HBM3, and an I/O speed of 5.2Gbps or higher. This is part of the company's "ambitions for even faster HBM solutions" than its current generation HBM2E memory chips. SK Hynix says its HBM3 memory solution is currently under development.
If the 665GB/s figure proves correct, then we're looking at a 44 percent jump in bandwidth compared to HBM2E. That is for a single stack of HBM3 DRAM, mind you. Take NVIDIA's A100 accelerator. The original model with 80GB of memory shipped with four stacks of HBM2, for 1.6TB/s of bandwidth. NVIDIA later released an 80GB version with five stacks of HBM2E, which increased the bandwidth to 2TB/s.
Five stacks of HBM3 would yield more than 3.3TB/s of memory bandwidth, based on SK Hynix's claim. And for a six-stack configurations (the 80GB A100 uses five out of six stacks), bandwidth balloons to nearly 4TB/s.
Bear in mind too that finalized hardware could be even better, depending on how development goes and where JEDEC lands with its official specifications for HBM3. Regardless, there is a ton of bandwidth at play here.
The caveat for consumers has always been cost. HBM is a bit more complex than GDDR, and the higher price is partly the reason why it is not commonly found on consumer graphics cards for gaming. There are some exceptions (AMD has dabbled with HBM and HBM2, as found on the Radeon VII and other GPUs), but they are comparatively few and far between.