SK Hynix highlights the huge size benefit of HBM over GDDR5 memory

for many years, laptop chips have become smaller and extra efficient via shrinking the dimensions of diverse features and locating methods to p.c. more transistors right into a smaller region of silicon. As die shrinks have end up greater tough, agencies have turned to three-D die stacking and technologies like HBM (excessive Bandwidth memory) to improve overall performance.

We’ve talked a amazing deal about HBM and HBM2 in the past few years, but photographic proof of the die savings is a piece more difficult to come by way of. SK Hynix helpfully had some HBM reminiscence on show at GTC this year, and Tweaktown caught photographic evidence of 8Gb of GDDR5 as compared with a 1 GB HBM stack and a 4GB HBM2 stack.

the one quibble i've with the Hynix display is that the labeling mixes GB and Gb. The HBM2 package is drastically larger than the HBM1 chip, however nonetheless a lot smaller than the 8Gb of GDDR5, notwithstanding packing 4x more reminiscence into its diminutive shape issue.

We don’t anticipate HBM2 to hit marketplace until the tail cease of this 12 months and the start of next; GDDR5 is predicted to have one closing hurrah with the launch of AMD’s Polaris this 12 months. those space financial savings, but, illustrate why each AMD and NV are moving to HBM2 on the high cease. Smaller dies suggest smaller GPUs with higher memory densities for each consumer, expert, and medical applications. technologies like GDDR5X, which depend upon 2nd planar silicon, can’t compete with the ability gain of layering more than one chips on pinnacle of every different and connecting them with TSVs (via silicon vias). GDDR5 will continue to be used for finances and midrange playing cards this technology, but HBM2 will likely replace it over the long time as prices fall, decrease-quit cards require more VRAM, and producer yields improve.

Over the long term, even though, even HBM2 isn’t sufficient to feed the desires of next-era exascale systems. The slide above is from an Nvidia presentation on excessive performance computing (HPC) and the strength requirements of DRAM subsystems. shifting to HBM drives a great improvement in I/O strength and an absolute development in general energy consumption for the DRAM subsystem. HBM2 draws less power to provide 1TB/s of bandwidth than GDDR5 used to show 200GB/s.

lamentably, sincere scaling of the HBM2 interface gained’t save you destiny reminiscence standards from exceeding GDDR5 electricity requirements. lengthy-term, extra upgrades and process node shrinks are nevertheless necessary — despite the fact that die-stacking has replaced planar silicon die shrinks because the number one overall performance motive force.